CharForge: rigged, animated game characters from a few words or one image, on a laptop

1 Introduction

A game character is more than a mesh. It needs a skeleton with the bone names an engine's animation library expects, skin weights that bend cloth without tearing it, hands with fingers, a face that can blink and speak, a collision capsule, levels of detail, and clips that move its feet without sliding. Image-to-3D models now produce convincing meshes from one picture [1, 2], and research systems produce rigged characters from an image or a prompt [3, 4, 5]; hosted tools and auto-riggers promise the rest [6, 7, 8, 9]. Two things are still hard. The first is the gap between a generated mesh and a character that deforms well: generated topology has no edge loops where joints bend [10], parts the generator fused (an arm to a hip, a vest to a sleeve) cannot move apart, and a rig built on them shows it in the first wave. The second is motion that is not in a library: a spell, a victory cheer, a particular roundhouse kick.

CharForge is our answer under a hard constraint: everything runs locally, on a 24 GB M5 MacBook, with openly available models. It is also an experiment in method. Each stage exists because a measurement showed the stage before it failing in a specific way, and each decision is recorded with the number that made it (Appendix A). This paper is the curated account of that record; the chronological version is the lab notebook.

Our contributions:

• An end-to-end local pipeline from a few words or one image to an engine-ready character (glTF and FBX, Mixamo bone names, LODs, capsule, face morphs, 18 clips) in about 41.3 minutes of median stage time (Section 3, Section 4).
• Generated-mesh repair for rigging: arms cut free of the bodies they were generated against, modelled hands sized to the forearm they end, garments bound to the bone they move with, loose pieces carried along (Section 4.4, Section 4.6).
• Moves from generated video onto the character's own rig: retrieval, a fit on the video's own body, and analysis-by-synthesis refinement of the skinned mesh, with hands and feet (Section 4.11).
• A measurement suite - silhouette and joint fidelity against the video, a per-stage breakdown that separates the rig's error from the scorer's, and synchronised four-view renders with joint readouts - and an account of the errors it found in itself (Section 5).
• A living, reproducible record: every table here is rebuilt from the pipeline's measurement files, and the experiment tracker is the project's roadmap (Section 9).

2 Related work

2.1 Characters from an image or a prompt

Image-to-3D generators. TRELLIS [11] and its successor TRELLIS.2 [1] generate meshes with physically based materials from structured sparse-voxel latents (a 1024³ asset in about 9.6K tokens). TRELLIS.2 conditions on DINOv3 image features [12]. Hunyuan3D 2.0 and 2.1 [13, 2] pair a shape diffusion transformer with a texture painter and produce PBR materials. Both families are general-purpose: they model a character as one surface, which is the root of most of what CharForge repairs. A practitioner's survey of hosted and open generators [10] recommends the same two-step workflow we use - draw a clean reference with an image model, then lift it to 3D - and notes the same weakness: generated topology lacks deformation loops at the joints. The Cinevva workbench advertises text or up to four views, rigging, animation and GLB export, but does not identify the underlying model [9]. AI3DGen advertises selectable-face-count OBJ exports and personal-use terms; its cited page does not document rigging or animation [14].

The Qwen-Image and Qwen-Image-2.0 papers describe earlier generations in the image-model family [15, 16]; the separate Qwen-Image 2.1 repository documents the checkpoint actually used by CharForge [17]. Its license permits non-commercial research and evaluation only; it requires a separate license for commercial use [18].

Character-specific systems. CharacterGen [3] canonicalises a character in any pose to four A-pose views with an image-conditioned multi-view diffusion model, reconstructs it with a transformer sparse-view model, and back-projects the views into a texture with a depth test and Poisson blending [19]; it rigs with AccuRIG [6]. Make-A-Character [4] takes the opposite route: a language model parses the prompt into attributes, a portrait is generated and fitted onto a fixed, production-rigged template (MetaHuman) using 431 dense facial landmarks, hair is built as strands, and garments are matched from a library by CLIP similarity. Make-A-Character 2 [5] starts from a single portrait, corrects its lighting and skin tone, and calibrates a facial skeleton for expressive animation. Template methods get rigging and faces for free but not the look of an arbitrary character; generation methods get the look but inherit the generator's topology. CharForge generates the character (like CharacterGen) and then builds the rig from the generated solid itself.

Decomposed and product-ready generators, and open peers. StdGEN [20] and StdGEN++ [21] generate a character as separate body, clothes and hair meshes from a canonical A-pose, the second adding a video-diffusion texture decomposition into editable layers; DreamCharacter-1 [22] post-trains a 3D foundation model for characters with geometry preference optimisation, two-stage texturing and distillation; and Pixal3D [23] conditions the TRELLIS.2 backbone by back-projecting image features along camera rays, so the shape is aligned pixel for pixel with the input view. Code is released for StdGEN and Pixal3D (CUDA; Pixal3D also through a third-party Apple Silicon port) and, as far as we could find, not for DreamCharacter-1 E153. Two open projects do the whole task on one machine: OpenMeshy [24] chains TRELLIS.2, Hunyuan3D and Pixal3D backends with retopology, a pass that copies the source picture's pixels onto the surface, SkinTokens rigging and Unity Humanoid export, and writes a record of the licences behind each output; reference-asset-compiler [25] gates every stage on recorded evidence, including a five-pose deformation suite for the rig, and never retries a crashed model run. The pages we read describe neither foot planting, interpenetration nor measured motion quality - the layers this paper adds. They had two things CharForge lacked, and took from them: a licence record with each package (PROVENANCE.json) and an extreme-pose check on skinning (Section 5.7).

Commercial generators. Meshy's API generates a mesh and texture from one picture, then optionally remeshes it to quad-dominant or triangle topology at a target count (100-300,000, default 30,000) and textures it at 2K, 4K or 8K with optional metallic, roughness and normal maps, delighting the base colour by default; its rigging step takes a textured humanoid facing +Z and returns a rigged model with walking and running clips [26]. Tripo adapts its face count to the shape (about 1.4 million faces, two million at its detailed setting), offers quad output up to 150,000 faces, a 'smart low-poly' topology of 500-20,000 faces, and texture tiers up to 8K [27]. Neither documents how its rig is weighted or how its animations are cleaned. Against them E162, CharForge's surface density is comparable (one million faces), but it ends in triangles, not quads. Its 4096-pixel texture holds about one megapixel of real detail, projected from an 832 x 1216 picture and 1024-pixel side and back views. Each clip carries the corrections we add to keep limbs out of the body. The roadmap takes these one lane each (Section 10).

Repair and representation quality. ManifoldPlus [28] turns triangle soups into watertight manifolds while optimizing the result back toward the input; Metro [29] measures simplified-surface error by sampling the source and output surfaces. These works motivate two separate checks in E116 and E118: topology repair success and geometric fidelity. They do not supply a CharForge-specific tolerance. Semantic-segmentation confidence is also not automatically correctness: Wang et al. [30] and de Jorge et al. [31] study calibration, reliability, and shift-related failure, motivating a manually labeled, held-out calibration set before multi-view part probabilities are trusted E117. Neither gives CharForge-specific thresholds.

Characters from drawings. Animated Drawings [32] rigs a child's drawing in 2D. DrawingSpinUp [33] lifts a drawing to 3D, removes and later restores its view-dependent contour lines, thins single-line limbs with a skeleton-guided deformation, rigs with Mixamo's auto-rigger from eight front-view keypoints and assigns each vertex to its nearest bone; its authors report surface adhesion where the drawing has self-occlusion, the same failure our arm cut addresses. Akman et al. [34] go from stick figures to 3D bodies with a variational autoencoder and animate by interpolating in its latent space.

Unified models. PixelUMM [35] reads and writes raw pixels with one decoder-only Transformer - images as patches, video as tubelets, no VAE or vision encoder - and both understands and generates images and video. Weighed against each CharForge stage E142, its released generation takes text only, at 256 px for an image and 176 x 320 for a video, where the stages need a full-body character at about 1024 px, image-to-image for the face and the character itself in its videos; its understanding scores are those of a 7-8B vision-language model, the level that could not judge moves from words here E143; and it runs only on Linux with CUDA. It is not used.

2.2 Rigging and skinning

Pinocchio [36] embeds a template skeleton in a mesh and computes weights by heat diffusion; RigNet [37] learns joints, connectivity and weights; UniRig [38] predicts skeletons autoregressively with a skeleton-tree tokenisation and was trained on over 14,000 rigged models. Commercial auto-riggers place a humanoid skeleton from guide markers - AccuRIG [6] with correctable markers and finger counts, PINOC [7] with mirrored markers from chin to ankles, Rig Lab [8] fully automatically, but with each hand a single joint, and Mixamo [39]. Magnenat-Thalmann et al. describe earlier joint-dependent local deformations for hand animation and grasping [40]. The glTF standard's linear-blend skinning path uses weighted joint transforms [41]; this can collapse bent joints, which dual quaternion skinning [42] can reduce, but not eliminate, by blending rotations as rotations. Our weights follow the idea of geodesic voxel binding [43] - distance measured through the voxelised body rather than through the air - so a vest glued to an arm is not near the arm's bone.

Learned rigging has moved on since: Puppeteer [44] adds attention-based skinning with topology-aware joint attention and an optimisation that animates the rig from video, and SkinTokens [45] encodes skinning weights as discrete tokens and predicts skeleton and skin as one sequence, reporting 98-133% better skinning accuracy than earlier methods. Both need an NVIDIA GPU (SkinTokens: 14 GB, CUDA 12.1 and flash-attn), so they are out of reach on this laptop; the geodesic binding we keep runs here, and the extreme-pose battery (Section 5.7) now measures it. Make-It-Animatable [46] is the learned skinning model that does run here: from a mesh it predicts a Mixamo skeleton, blend weights and a rest-pose reset with a transformer trained on Mixamo characters, and of its CUDA requirements it uses only farthest point sampling and PyTorch3D's transforms, both plain PyTorch - 48 seconds a character on a CPU. Its weights on our skeleton are measured in Section 5.7 E157. AniGen [47] generates shape, skeleton and skin together from one image, the route that would remove the fused surface's rigging problems at the root; a third party's run on an Apple laptop rigged a simple mascot and produced no faces for a clothed character with thin limbs, so it is not tried. For faces, OmniFaceRig [48] registers a template to the character's dense face keypoints, builds the inner mouth procedurally (teeth, gums and tongue, their texture baked) and transfers up to 155 FACS shapes with collision awareness, on an A100 in 20-30 seconds; the template route of E149 follows the same recipe with the Anny body's 52 ARKit-named facial actions [49]. Single-photo face reconstruction has a strong 2026 prior in Pixel3DMM [50] (per-pixel normals and UV from DINO features constraining a FLAME fit), which needs CUDA and a non-commercial licence.

Expressive hands and faces are often modelled together with the body: SMPL-X [51] extends a body model with articulated hands and an expressive face, while MANO [52] learns pose-dependent hand shape corrections from scans. These works motivate testing hand and face deformation as well as landmark placement; their research models are not drop-in templates for arbitrary generated characters. InterHand2.6M [53] provides real-captured 3D hand-pose labels, while D3DFACS [54] and the FaMoS sequences released with TEMPEH [55] provide dynamic human-face/action evidence. Their different topology, style and access terms mean they are calibration references, not direct acceptance tests for stylized meshes E123 E124. AU-Blendshape [56] specifically studies stylized action-unit controls and provides a relevant challenge-design reference, but not a CharForge benchmark or a licensed drop-in asset set. Pose-space deformation [57] provides a classical basis for the elbow and shoulder correctives planned in E107.

Layers and motion. CAPE [58] learns pose-conditioned clothing shape, and CLOTH3D++ [59] pairs animated dressed-human sequences with dense cloth geometry and texture. They provide useful challenge cases for layered motion, but neither demonstrates CharForge's runtime spring chains or guarantees stylized garment behavior E110 E125.

2.3 Motion from video

Human pose models read 2D joints from images: ViTPose [60] for the body, DWPose [61] for the whole body including 21 points a hand. Recovering 3D motion from them is usually done by fitting a parametric body such as SMPL [62] - by optimisation, as in SMPLify [63], or by regression, as in HMR, 4D Humans and WHAM [64, 65, 66], the last grounded in the world with foot contact. We do not fit a parametric body: the target is the character's own skeleton and mesh, whose proportions (a cartoon's head, an anime character's legs) no human body model shares. Instead we retrieve the nearest clip from the 2,453-animation Mixamo mirror [67] by dynamic time warping [68] of its 2D projections, bend it toward the video, and refine the character's own skinned mesh against the video in the spirit of SMPLify. The videos themselves come from the community Ref2VA fine-tune [69] based on MiniMax H3 [70], conditioned on the character's picture; Comfy-Org supplies separately packaged encoder and VAE components [71].

Motion ground truth. AMASS [72] unifies optical mocap datasets as rigged 3D motion, making it useful for controlled renders with exact motion truth but not paired RGB. TotalCapture [73] provides synchronized calibrated video, IMU and Vicon labels; 3DPW [74] adds in-the-wild footage from a moving phone camera with IMU/video-derived poses. AIST++ [75] broadens action coverage with multi-view dance footage and reconstructed 3D motion, but its annotations are not Vicon measurements. These sources test different failure modes and label quality, so we keep them as separate conditions rather than pooling them E111.

Contact labels. UnderPressure [76] records pressure insoles with motion capture and derives foot-contact labels. It supports evaluating contact classification against an independent signal; the geometric sole-depth threshold and foot-slip ratio in CharForge remain separate proxies and need their own calibration E122. Pražák et al. [77] found small footsliding perceptible in most tested conditions, but their controlled mannequin study does not establish a universal acceptance threshold for generated characters.

2.4 Measuring generated assets and delivery

TIFA [78] and GenEval [79] use fine-grained question or object checks to expose prompt attributes, counts and relations that a holistic image-text score can miss; neither is validated for stylized character renders. MATE-3D [80] instead collects human ratings for semantic alignment, geometry, texture and overall quality separately. THUman 2.0 [81] supplies scan meshes and textures for a permission-checked scan-anchored study; rendered orbit views are synthetic views, not captured calibrated camera observations, and redistribution of derived data requires written permission under the posted terms. We use these works to design and calibrate character-specific measurements, not to assume a published image or 3D evaluator transfers unchanged E102 E112 E113.

Metro-style surface distance [29] complements the existing nonmanifold checks with source-to-retopo shape error E118. Texture Mapping Progressive Meshes [82] motivates checking local UV stretch and anisotropy beyond median texel density E119. Unity's mip documentation [83] describes runtime selection and trilinear blending across downsampled levels; therefore a base-resolution gutter alone does not establish that exported color, normal and ORM maps are seam-free at runtime E120.

For animated humans, DDH-QA [84] includes human-rated binding errors and motion unnaturalness, while DHQA-4D [85] studies geometry, texture and temporal-discontinuity distortions in scanned 4D humans. These taxonomies are useful for designing CharForge's deformation audit, but their learned quality predictors are not assumed to generalize to generated, stylized characters E102.

The Khronos glTF Validator [86] checks conformance to glTF 2.0, while Unity, Unreal and Godot document importer-specific behavior for rigs, clips, morphs, materials and skin influences [87, 88, 89]. A standards-valid file is therefore necessary but not sufficient evidence of engine-ready behavior; the importer/runtime smoke matrix is E114.

For traceability, W3C PROV-DM [90] models entities, activities, agents and derivations; CycloneDX's ML-BOM [91] gives an inventory pattern for models and data, and Sandve et al. [92] emphasize archiving exact computational procedures. They motivate a claim ledger and output-lineage manifest, not a claim of formal provenance compliance E115 E121.

3 System overview

CharForge has three lanes that share one skeleton (Figure 2). The character lane (charforge.py make, 23 stages) builds a character once, from a prompt or an image. The move lane (charforge.py move) adds a move from a generated video to any character; because every clip it produces is on the standard Mixamo skeleton, one video serves every character. The words lane (tools/unimate_moves.py) adds a move from a sentence, generated on the character's own skeleton by UniMate [93] E129. Both lanes are Python and headless Blender [94] around local models: an image model for the reference [17], TRELLIS.2 [1] through its MLX port [95] for the 3D model, a human parser [96, 97] for part labels, ViTPose [60] and DWPose [61] for pose, U²-Net [98] for cut-outs, a small language model [99] for prompts, and the community Ref2VA fine-tune [69] of MiniMax H3 [70], with separately packaged Comfy-Org encoder/VAE components [71], through ComfyUI [100]. A browser application, the Studio, runs the same stages for a user without the command line (Figure 19).

Where the time goes. Table 1 gives the median time of each stage over every build logged on this machine. Four stages - the second, multi-view 3D pass (off by default since E106), the reference image, the first 3D pass and the texture projection with its detailed face - take 37 of the 41.3 minutes; everything that turns the generated mesh into a rigged, animated character takes under five. The times include any wait for the GPU while another job held it (Section 4.13), so they are what a user waited, not each stage's own cost; timing each stage alone, and cutting what the measurements say is not worth its time, is experiment E106.

4 Method

We describe each stage by what it does and the failure that made it necessary. Experiment identifiers link to the tracker, where each has its question, metric and result.

4.1 A few words, written out

Users type little: "a boy scout", "gray alien that is extremely muscular". Passed straight to an image model, the first came back a grown man at the height slider's default of 1.72 m and the second a grey clay sculpture - no clothes, no colour. Following Make-A-Character's use of a language model to parse a prompt into attributes [4], a local model [99] fills a fixed set of fields - who, age, build, skin, hair, face, top, bottom, shoes, extras - and the sentence is composed from them in code. Asked instead for free prose under rules, 3B-8B models broke a rule in about one prompt in two ("holding a wooden badge", "standing at 1.5 meters tall"); fields that name a prop, a pose or a height are now dropped, and the age group is taken from the prompt's own words, not the model's. The height follows the age group unless the user sets it, and a name that exists is refused rather than overwritten E016.

We scored nine local models on five prompts against four rules each E015; Table 2 gives the results. Gemma 2 9B kept all 20 checks but took 15.0 s a prompt; Gemma 4 e4b kept 19 in 9.2 s and is the default. Gemma 4 answered nothing until its "thinking" mode was switched off, which only models that report the capability are sent. The Studio shows the written-out text before anything is drawn, to edit.

4.2 The reference image

Everything is generated from one picture of the character standing in an A-pose on a plain background. We compared two local image models on six characters with the same prompt and seed E013: Krea 2 drew one broken figure in six (a jacket with no head or legs, which the automatic checks passed, since it still has "arms"), Qwen-Image 2.1 [17] produced no broken figures and returned a usable alpha cut-out in each case. Qwen read "arms slightly away from the body" literally - wrists 0.28-0.37 torso lengths from the hips, close enough to fuse in 3D - until the angle was spelled out (0.58-0.70). It is the default; Krea stays a flag away. Qwen-Image 2.1's license is limited to non-commercial research and evaluation; commercial use requires a separate license [18].

A reference in the wrong pose costs half an hour before it shows. Bo, a cartoon chef, was drawn with his fists on his hips despite the prompt; the 3D model fused them and his hands came out in shreds at his belt. The pose model now reads every reference's arms: our references put the wrists 0.60-0.79 torso lengths from the hips, Bo's 0.33, and below 0.45 the picture is drawn again on a new seed E028.

4.3 The 3D model, checked from the front

TRELLIS.2 [1] builds a mesh with PBR textures from the reference. On four characters its first pass built the picture's plain backdrop into the model as a solid board - in front of the figure, behind it, or through its legs - which later stages cut into grey sheets hanging from the arms (Figure 3). Rendered from the front, a sound model lies on the picture's silhouette at IoU 0.9-0.94; these lay at 0.35-0.44. The model is now checked from the front, and below 0.8 made again from the picture cut out by a salient-object model [98], then on a new seed E014: Gray 0.42 → 0.931, Knight 0.4 → 0.916, Kaito 0.35 → 0.906, the boy scout 0.35 → 0.908. It is also checked from behind: on Vex, made from a painting, TRELLIS.2 sculpted and painted a second face on the back of her head, which nothing later looked at. DWPose [61] reads a face on the back of an ordinary head at 0.34-0.55 - a ghost of one in any hair - and on hers at 0.97; above 0.8 the model is made again on another seed E145.

Pixal3D, on request E155. Pixal3D [23] conditions the same backbone on image features carried along the camera's rays, so the shape lines up with the picture. Run through its Apple Silicon port on six of our references with the same gates, it matches the picture's front silhouette better on all six (IoU 0.921 → 0.949 on average) and comes out as one closed piece with a quieter surface and clean armour and cloth edges, where TRELLIS.2's first pass is 40-409 pieces. It is worse where the gates do not look: the mirror symmetry of five of six, black eyes on two, a face on the back of the head read higher on two, and two to three times the time. So it is an option (--generator pixal3d), turned to face our cameras and welded at the import's seams before it is decimated, and TRELLIS.2 stays the default.

A second pass conditions TRELLIS on side and back views repainted from renders of the first. Scored against a generated 360° video of Mara E012, the second pass moved the mean silhouette IoU from 0.778 to 0.781, and three turntable frames as extra views to 0.786: more views barely change the shape, and where views disagree the model averages them (Mara's braid, over her shoulder in the reference and down her back in the video, was lost at the back). The second pass is also the slowest stage (Table 1). Every character built with it keeps both models, so the two were held against each other on 12 characters E106 (Figure 4). The second model fits the reference's front worse on 12 of 12 (silhouette IoU 0.923 on average for the first pass, 0.908 for the second), and its side silhouettes overlap the first's by only 0.63-0.93: it is a different model, not a refinement. Where the difference shows it follows the repaint, not the picture. The image model repainting the side and back views invents what the reference does not show, and the second pass builds it: the boy scout grew knee socks and lost his grin, Vex's trainers became boots and a grey hood appeared under her jacket, Mara's braid moved from her shoulder to her back. What it adds is a more plausible back (the scout's neckerchief is tied behind him). The second pass is therefore off by default since this version (--quality best keeps it), which saves a median 16.1 of 41.3 minutes. This is a pilot on finished models: the end-to-end comparison, with the characters rigged and audited both ways and rated blind, is still to run.

A finer surface. TRELLIS.2 simplifies its surface before writing it. At the 200,000 faces this pipeline asked for until version 0.11, gently curved armour became large flat facets with sawtooth edges, and every later stage kept them: the solid's vertices are put back on the generated surface to keep its detail, so the facets reached the normal map and showed as a triangle pattern under any grazing light (Figure 5). The generator now keeps 1M faces, and its surface is filtered before anything is built on it by bilateral normal filtering [101]: each face normal is averaged with its neighbours', weighted by their area, their distance and how alike the two normals are, so facets a few degrees apart become one smooth surface while a plate's edge, tens of degrees sharp, stays; the vertices are then moved to agree with the filtered normals [102], a median 0.05 mm on a 1M-face model E135. A rebuild of the roster on it is scored in Section 5.8. Quad remeshing was tried as well and refused both the generated mesh and the solid (QuadriFlow needs a manifold input).

4.4 Solid, joints, arms and hands

The generated mesh is closed into one solid (OpenVDB level sets [103]), its flakes, inner shells and shards removed. Orbit renders give per-vertex part labels from a SegFormer architecture [96] using the FASHN human-parsing checkpoint [97] and 3D joints from the pose model; each joint is then moved onto the centre line of its limb, traced through the solid. A traced line can double back: on knight2 the arm's trace reached his gauntlet and turned 36 cm back up the vambrace, and the hand was placed a forearm's length behind the wrist; the trace now ends at its farthest point E018.

Arms cut free. Generators glue arms to whatever they hang against - a hip, a vest, a bag. Rigged as is, the surface between moves with both, and a raised arm pulls the vest into wings. The arms are cut free below the armpit (T-pose faces stretched past 3x: 1,150 → 390 on Pip), thin webs opened first, and the cut checked slab by slab along the arm and repeated where a join survives: 3-9 joined slabs an arm after the first cut, 1-3 after the second E025 E026. The cut is a shell at the arm's estimated radius, and at the armpit the estimate runs large: on 8 of 17 characters the shell went through the torso from front to back and showed on the back as dark rectangular holes. Once the arm is free, everything the cut took that lies more than 3 voxels (about 5 mm) from the arm's own solid now goes back to the body; a band of air still parts them wherever they were glued, and the stage undoes the give-back if its own joined test would fail E146. That band is also where the black squiggles on Rowan's jacket come from: seen from behind, its mouth runs up the back of the shoulder as a slot a few millimetres wide that light cannot reach (the baked occlusion there is 11 of 255), and no change to the paint shows through it. Widening the band to 6 mm halves the slot's vertices but leaves the marks as they were and lets the arms go deeper in the clips, so it is not adopted; the slot's mouth has to be closed above the armpit E169.

Modelled hands. Generated hands are paddles. They are carved off at the wrist and replaced by a modelled hand (a signed distance function with anatomical proportions, 15 finger bones) unioned onto the solid. The modelled wrist was a person's; on Gray, "extremely muscular", whose forearms measure 11-13 cm at the wrist, each arm ended in a flat step like a cuff and the waving hand read as a small hook (Figure 6). A sleeve's cuff measures just as wide, so width cannot tell a bulky forearm from a sleeve - but the part labels can: the vertices just above the cut are labelled bare skin on 0.92-1.00 of a bare forearm and 0.13 or less under a sleeve or gauntlet. On a bare forearm the hand is now scaled across its width and thickness to meet it, up to 1.6x (Gray 1.60x) E021.

The exact boolean that joins a hand to the arm fails now and then, and the fallback was to join the hand as an overlapping shell. On five characters the left hand fell back, and the right hand's boolean, run on that self-intersecting body, returned it unchanged; the only check, that the volume did not shrink, passed it, and they shipped without a right hand. Every hand is now unioned onto the body as it came from the solid, and a failed one joined only after all the unions; a union is kept only if the volume grows by at least a quarter of the hand's, four or more fingertips lie on the result's surface and it adds few non-manifold edges (exact, then nudged a tenth of a millimetre, then with self-intersection handling, which put hands on but tore the surface). The stage fails if a hand's fingertips are not on the final mesh, and after retopology every fingertip the hand model recorded is looked for on the low-poly E146.

4.5 Mesh and texture

The solid is decimated to about 60,000 triangles [104], the head keeping 28% of them, unwrapped, and its colour, normal, roughness/metal and ambient occlusion baked from the generated model. Collapse decimation spends triangles where the surface turns and left the head 17% of the budget, too few for an eyelid or a lip; a vertex group lets the head's vertices collapse less freely, its strength found by bisection until the share holds E140. The generator's own colour is a 1024-px atlas of thousands of small charts for a million faces - most triangles smaller than a texel - and along the charts' edges it leaves texels sampled off the surface; baked, they were a crackle of short grey and dark lines over every surface the picture does not paint (Figure 7). They are not dark specks a filter on the atlas can find, nor holes in the mesh; and an atlas texel cannot be mended from its neighbours there, which belong to other charts. So the colour is rebuilt through the mesh: each triangle sampled at its centroid, each vertex (its copies at the atlas's seams joined) given the median of its triangles' colours, a vertex far from its neighbours' median along the mesh given that median, and the atlas painted again from the vertex colours at four times its resolution, where nearly every triangle holds a texel of its own. New characters are generated at 2048 px. Smart UV projection on a smoothed copy folded islands onto themselves wherever the copy curls (224-2602 faces per character sharing texels with another surface - the backs of Pip's sleeves showed his vest's paint); the overlapping faces are unwrapped again as their own islands at one texel density, leaving 6-31 E027.

The bake is soft, so the reference is projected back onto the mesh, aligned to renders of the mesh by a similarity transform on the silhouettes and then by a smooth optical flow [105], together with the repainted side and back views; the reference sets the palette, colour-matched group by group. Like CharacterGen's back-projection [3] a view only paints surfaces it sees face-on and unoccluded; unlike it, views are blended by weight rather than by Poisson blending [19] - comparing the two is E105. The face is painted from a detailed image: the reference's face region enlarged 3-5x and re-detailed by the image model at low strength.

Sharper from the same pixels E164. The 4096 atlas holds about one megapixel of real detail: the reference is 832×1216 and the views 1024 px. Three changes take more of it into the texture. The pictures are read with a Lanczos filter instead of bilinear. The bake's own colour, which was mixed in at 8% everywhere, drops to 0.1% where the views see a texel strongly, and keeps only its low frequencies, as the colour field the views' detail is laid on. The shadow floor below was a softplus with a slope of 0.2 at black, which flattened the darkest folds; it is now a power curve whose slope never falls below 0.4. On Mara, Pip and Hana the albedo's lightness detail rises 14%, 13% and 3%, and an unlit front render keeps 0.51, 0.58 and 0.65 of the picture's own detail at the picture's scale, up from 0.43, 0.52 and 0.59. The sides and backs are limited by their 1024 px views; a 2× refine of the reference and views (pipeline/tex_hires.py) is written for when ComfyUI runs.

The face in parts. Faces looked sloppy after version 0.12 - eyes hollow or smeared, features off the shape, patches on the cheek - and taken one part at a time the faults had separate causes E140. The eyes were in the shape: TRELLIS.2 models every eye as an almond opening with shards inside it, the solid filled it with streaks, and the decimation crumpled it into a pit no texture reads as an eye (Figure 8). Each opening is found by DWPose [61] on a front render of the generated model in its own colours - the generator paints its features where its geometry has them, where on a grey render of a drawn face the pose model only guessed - and rebuilt in the solid as a cap across the opening: it follows the lids' depth round the rim, bulges a little forward in the middle like an eyeball under lids and sinks a little at the rim; what lies in front of it is cleared, the region smoothed and kept from being put back onto the generated surface's shards. A ball set behind the opening instead left a slot whose hard ledges rendered as dark rings round the eyes. The features were in the projection: the silhouette fit had put Mara's painted face larger than her sculpted one, its eyes 10-30 pixels off. Each part - each brow, each eye, the nose, the mouth, the jaw - is now moved by its own similarity [106] from where the projection puts the model's points (read on the generator's paint) to the picture's, and the moves are spread by a thin-plate spline [107] that falls to nothing past the face; this replaces the earlier face flow, which had squinted and smeared Mara's eyes. The face's ellipse is painted by the front view alone - the generated side views, their faces drawn again a little elsewhere, put patches on the cheek and skin on the hair - and left out of the skin tone below (Figure 9). Anime faces are not laid on in parts: a drawn face's shape does not say how big its eyes are, only the generator's paint does, and fitted to it Kaito's eyes came out a fifth wider than his picture's. Where DWPose cannot read the model's eyes (12 benchmark characters: 11 rebuilt; most anime faces are not read) the eyes stay as generated, and where it cannot read both faces the face flow is used as before. A drawn smile can run past the lips - Cadet's, Kaito's and Bo's corners read 1.2 eye spacings apart - so the jaw's cut is held to 0.8 of the eye spacing.

No loose shards. The solid is one piece, but the modelled hands are unioned onto it and the decimation's repairs delete faces round non-manifold edges, so small shells survived into every low-poly: 2-94 a character, of 1-58 faces each and a few millimetres to 3 cm across, floating at Cadet's and Knight2's shoulders, and holding nearly all of the model's open edges (Cadet 207 of 212). Pieces under 0.25% of the faces are dropped; a garment or a strand kept apart by design is hundreds of faces or more E138.

Hands shaded by their own shape. The modelled hands take the skin tone of the face, and took their light and dark from the base colour bake - which comes from the generated mesh, where the old hand, or nothing, stands where the modelled hand is - so every hand carried pale and dark patches. They now take it from the ambient occlusion baked on the solid they are part of: soft shade in the creases between the fingers, nothing elsewhere (Figure 10) E138.

One skin. The skin's colour comes from three places - the picture, with its light and shadow, on the front; the generator's colours on the sides and back; one tone on the modelled hands - so the same skin was three colours: a ring at every bare wrist (Mara's forearm L* 50 against her hand's 69, CIEDE2000 16.6) and, on the face, the picture's shading lying on a shape not quite the picture's. Skin texels - near the character's skin tone in hue and chroma, at a lightness within a band of it so that shading counts and brows, lashes, lips and eyes do not, on the parts the parser calls skin (brown boots and khaki are skin-coloured too) - have their colour at the 1-2 cm scale of shading replaced by the tone, measured in 3D so a cheek's islands are not kept apart by the atlas, and keep their own detail at the scale of freckles and pores. Mara's wrists go from dE 16.6 and 16.3 to 1.5 and 2.2 E138. Anime is left as drawn: its painted shading is the style. So is the face, since version 0.13: its colour bands cut the picture's soft cheek shadow into patches, some texels lifted and their neighbours not, and the face is now the picture's own E140.

Colours a renderer can light. A physically based renderer multiplies the light by the base colour, and the PBR guides keep a non-metal between sRGB 30-50 and 240 [108]: a black jacket painted at sRGB 5-15 renders as a hole without folds, and 17-38% of five characters' textures lay below 30. Every non-metal texel's lightness is eased into that range by a smooth curve that leaves texels well inside it untouched and keeps a garment's own contrast, its hue and chroma as painted; metals (the ORM map's metal channel) are left alone E132.

Painted sides and backs, on request. Hunyuan3D-Paint 2.1 [2], run through its Apple Silicon port [109], paints six views of the mesh from its own normals and positions, consistent with the picture and each other. Its own bake back onto the atlas covered a third of the surface on our meshes and filled the rest in UV space, a patchwork, so only its views are kept, fitted onto our cameras by silhouette and projected like any view, with the front delit by Hunyuan's delighting model and each garment's colour on the side and back views matched to the front's E133 E134. It cleaned Pip's blotchy back. On eight more characters, though, the texture audit scored it better while it looked worse on six: generated at 256 px, the views paint what the picture does not show - a face-like pattern on the back of the boy scout's head - and lose fine pattern (Figure 11). It is an option (--texture-views hunyuan), used for Pip.

4.6 Weights and rig

Skin weights come from distances measured through the solid body rather than through the air [43], so a vest that touches an arm is still far from the arm's bone. Three rules handle garments the generator fused to the body E025. The arm's own surface is flooded from the sleeve's outside starting past the cut (0.30 of the upper arm), so it does not walk onto the vest; a vest's armhole moves with the collarbone where colour tells the torso garment from the sleeve; and an upper garment near the hips follows the pelvis rather than the thighs, so a jump does not pull a hem into a W. Loose pieces the generator left (Kaito came as 40: hair spikes, the inner toes of his sandals) each ride rigidly with the body part nearest them. The project rig maps the 64-bone Mixamo-style vocabulary in its clip library, including fingers and a jaw [67]; bone chains are hung on what should swing (braids, ponytails, bags). The glTF standard's skinning path is linear blend [41]; the playground patches three.js [110] to dual quaternion skinning [42], which cuts jacket faces collapsed below half their area during the walk from 8.8% to 6.1% and during the run from 11.5% to 6% with the same weights E024.

Two weightings, and the extreme poses choose. A gentler falloff (influence by distance to the power -2 instead of -4) and more smoothing (30 passes instead of 6) take the stretch out of a loose jacket's armpit: with Vex's arms overhead, stretched, sheared and crushed faces fall from 13.3% to 10.5%, with her arms forward from 10.4% to 7.1% E154. On the legs the same weights cost the strides: chosen for nine characters in a first rebuild, they stretched Juno's and Mara's thighs more in the gaits and took Knight from 2 pops to 6, five of them in the sprint; on the head and neck they stretched Juno's and Knight's heads in the clips. So the soft weights are blended with the default by how much of a vertex follows a hip, knee, ankle, the neck or the head - soft on the body and arms, the default on the legs and head - and each build makes both, rigs each through the same stages and poses it in the ten extreme poses of Section 5.7. The hybrid is kept when its mean share is at least 0.1 points lower and no pose is worse by more than 0.3 (Table 11): 10 of 16 characters keep it, and the tight-clothed and armoured ones keep the default (Ren's deep squat loses 0.35 points with the hybrid). Knight rebuilt with the hybrid is back to 2 pops; keeping the head on the default too took Juno's stretched faces in the clips from 1.46% to 1.40% and sheared from 1.65% to 1.59% E160.

What the extreme poses cannot judge. Two other weightings scored better still at the extreme poses and worse in the clips. Make-It-Animatable's learned weights [46] beat the hybrid on six of seven characters - on Vex by 15% - and broke Cadet's pauldrons; taken through Vex's clips they stretched and sheared less and left a part more than 2 cm inside another in 22% of frames against 7% for the build it replaced, with 8 pops against 3 E157. Measuring distance from each body part's surface instead of its bone's line, so that a loose jacket's side - which hangs nearer the thin arm's bone than the thick trunk's - stays with the trunk, takes the fin out of Vex's raised arms (Figure 13; the fin vertices' arm weight falls from a half to a twentieth) and cuts her extreme-pose share from 4.35% to 2.73%; in her clips, frames with a part inside another went from 7% to 12% and pops from 3 to 15, the arm now sinking into a jacket that no longer moves with it - while Pip's clips, his vest no longer rising with his arms, improved on every measure (deep frames 6.8% to 4.6%, stretched faces down a fifth, a pop fewer) E159. Faces that hold their shape can be faces that let a limb through them, and the battery counts only the first (its own penetration reads the same for Pip's and Vex's two weightings); both stay experiments (--skin-weights mia, geodesic_weights.py --radius) until the clips can decide. Now they can: tools/weights_trial.py builds a candidate through every clip in a copy of the character and keeps it only when no clip is worse beyond a tolerance E161. The surface-relative weights fail it on Vex, Pip, Rowan and Bo - Pip's clips stretch and shear less, but his fall and strafe go deeper - and the cause is in the method: the radius is subtracted per bone, so a thick bone takes the weight of the thin joint beside it on the same chain (Vex's left wrist keeps 33 of its 880 vertices; her forearm and hand pass through each other), and the trunk's radius pulls weight from the shoulders and collar. The same judge chose Pip's weights for 0.20 E171. The colour rules read the texture, and 0.19's sharper one moved 1.8% of his clothing vertices - enough to tip the battery from the hybrid to the default by 0.005 under its margin. Of the variants tried through his clips, soft weights on the arm-bound vertices only, the default everywhere else, take his frames with a part more than 2 cm inside another from 6.4% to 3.9% with every other measure better; its one worse clip is one more of sprint's eleven sampled frames, and a clip's tolerance is now at least one sample, the audit's own resolution.

The pose it was modelled in, as its rest. Retargeting turns each source bone by its change from its own rest and applies that change to the target's rest, so the two rests must agree: Mixamo's is a T. Until version 0.11 the rig was therefore skinned up from the A-pose the character was generated in to a T and saved in it - and skinning is linear blending, which cannot be undone exactly: armour fused to the arm (Cadet's pauldrons) sheared into shards on the way up, and arms-down frames, skinned back down from that T, kept slivers. The T is now recorded on the armature and used only inside the retargeting maths; the rest pose stays the modelled one. The clips come out the same - foot slip identical in all nine gaits - while on Cadet crushed faces went from 3.44% to 1.446%, self-intersecting face pairs per frame from 1,093 to 502, and frames with a limb more than 2 cm inside another from 40% to 28% (Figure 14) E137. Arms raised above the shoulder still stretch armour fused to the arm; that needs rigid shoulder pieces E110. Engines take an A-pose bind like any other (Unity's Enforce T-Pose, Unreal's retarget pose, Godot's Fix Silhouette); --rest T bakes the T as before, and the manifest says which a build has.

4.7 Face rig

The head is rendered from the front, unlit, together with the world position under every pixel. Eyes are found as blobs and the mouth by its colour and the pose model's face points; from them the face stage cuts the mouth open, adds a jaw bone and five shapes (blinks, smile, raised brows, pucker), and beside them 23 shapes under ARKit's names - eyes wide and squinting, brows down and raised inside and out, jawOpen, smiles, frowns, funnel, stretches, cheek puff, sneers - the names engines and face-capture apps drive E166. Each is a smooth field per side from the same landmarks, gated by the same trust in the eyes and mouth; jawOpen is the jaw bone's opening as a morph. Stored as sparse morphs, they add about 0.6 MB to a web model. The mouth is cut along the seam; where the plane through the seam and the corners tilts more than 10° while the seam itself sits level with them - a mouth seen a little from the side - the cut is made level, so the jaw does not open along a slant (Knight's folds on jawOpen 121 → 95, the pouch from nose to cheek gone) E170. It places nothing it cannot find: on Bo, eyes behind round glasses and a mouth under a moustache, it builds no face rig rather than a guessed one E109. Two characters lost their face rig because the frame stopped above the chin on a tall head E017. The position render had its own error: written as position + 4 into EEVEE's half-float film, every position was read in 3.9 mm steps, 2.5 mm off on average; written around the frame's centre the step is 0.5 mm E019.

4.8 Clips, feet and packaging

Eighteen motion captures (idle, walk, jog, run, sprint, backpedals, strafes, turns, crouch, jump, fall, land, wave) are retargeted onto the rig with the fingers. Feet are planted by clip selection, grounding and foot IK, and the playground drives all locomotion from one stride clock, so the foot on the floor slides a median 3.2% of the gait's ground speed (Section 5.5) E029. Every change of clip is a cross-fade (0.12 s), and the simulation steps at no more than 1/120 s whatever the display's rate, so a gait change, a landing or a wave no longer jumps in one frame (Mara's worst single-frame bone turn: 95° → 43° between gaits, 136° → 24° on landing) and the motion is the same at 60 and 120 Hz E167. The foot that is planted is the shoe as the viewer deforms it: every vertex of the shoe below the ankle (1016-3266 a foot on the roster), each posed by its own skin weights with dual quaternions, the way the playground skins the mesh E130. A planted key pivots on the flat of the sole; a key in the air keeps the whole shoe, toe cap included, above the floor. A rigid copy carried on the ankle bone works for an ordinary shoe, which is rigid below the ankle; Cadet's armoured boots blend with the shin above a high ankle joint, and the rigid copy stood on the floor while the boot on screen was 4.4 cm under it. All of the shoe is needed, not a sample of it: thinned to 400 points, its lowest point jumped from one sample to the next as the foot rolled, and the planted pivot jumped with it. The package follows engine conventions: real height in metres, origin on the floor under the pelvis, glTF [41] and FBX with three levels of detail, and a manifest with the collision capsule, each clip's ground speed, direction and contact phase, the jump's take-off and landing, and the face's morph targets.

4.9 Room for each limb

A motion capture was performed by one person of one build. Played on a thicker character, the same joint angles fold a forearm into a puffy sleeve, a heel into a thigh, and hang the arms inside a wide torso. Game rigs carry joint limits for this; we measure them for each character from its own mesh E128. From the rest pose each elbow and knee is bent in 5° steps about its hinge axis, read from the character's own clips as the mean axis of each frame's bend, and the upper arm is lowered from the T-pose at five swings, from 60° forward to 60° back. At every step the mesh is posed and the depth of one body region inside another measured as in Section 5.6. A joint's limit is the last angle before that depth passes 2 cm, the audit's own threshold for a deep penetration: flesh squashes where a rigid mesh can only overlap, and at a 1 cm tolerance Mara's boxing guard was opened by 25°. The arms' limits leave out the head, whose region includes the hair; hair should swing aside, not stop an arm.

Every clip is then held inside the limits. An arm past its limit moves to the nearest allowed direction, measured against the chest so a lean does not count, and those turns are smoothed over ±2 keys so the nearest choice cannot hop between two near-equal ones (on Bo's first limits, raising the arm instead cost up to 44° where moving it out to the side cost 17°), and faded out as the arm swings past 60° forward or back, where the limits were not measured. An elbow or knee past its limit is opened about its own bend axis, a knee only while its foot is off the floor (faded in between 8 and 16 cm of lift), so the feet stay where retargeting planted them. The limits are soft: the last 10° are approached along a tanh, so a correction has no corner in time. The sweeps take about a minute per character and the clearance seconds; both run at the end of the animate stage. Contact-aware retargeting treats the same interpenetration with a learned, geometry-conditioned model [111]; measured limits need no training data and cannot invent a pose, but they only see each joint alone (see Section 5.6 for what that leaves).

4.10 Every frame against the body

Limits are measured with each limb beside the body and held one joint at a time, so they cannot see a hand go into a thigh that is itself swinging, or a forearm into a bulky jacket half-way through a turn - after clearance, those pairs still carried most deep frames E148. The research answer is retargeting that models the meshes' interaction: MeshRet [112] learns a dense interaction field between body surfaces and runs on this laptop's CPU in under a second a clip. Measured on six characters E150, it removes arms from torsos but drives the hands into the thighs, a pair its field does not include, and its median share of deep frames (35%) is no better than clearance's (36%). We keep its idea and drop the network E151. After clearance, each clip is solved against the character's own skinned surface: for every frame, the smallest and smoothest extra turn of each shoulder, elbow and wrist (in the bone's own frame; the fingers ride the wrist) that keeps about 1,500 points of each arm's surface 1 cm outside the torso, thighs and shins, posed from the clip by linear skinning. A point counts as inside when it lies behind the normal of its nearest body vertex and within 2.5 cm of that normal's line, and the test runs both ways round: Pip's puffy vest goes through his sleeves, which only a body vertex measured against the arm's surface shows. Points with more than a tenth of their weight on the other part or a collar, or within 3 cm of it at rest, are not tested - that is the fold at the armpit, where arm and torso are one surface - mirroring the audit's own rule. The cost is the squared depth of every inside point, plus the squared turns (a twist three times, an elbow's sideways turn five times dearer) and twenty times their squared change from frame to frame, minimised over the whole clip at once (Adam, 300 steps, nearest vertices re-found every 25). A clip with nothing near a surface is left as it was. The turns are written into the clip's own keys, so the planted feet and the pelvis are untouched. Contact-aware retargeting [111] optimises in a learned space for the same reason; here the space is the joints' own turns and the measure is the mesh.

Three refinements, each from a residual the first version left E152. The cost's squared depth is weighted ten times the squared turns: at equal weight, when only a handful of points were inside, the turn penalty outweighed them and Gray's idle stayed at 98 of 98 deep frames although the same solve cleared it at a wider margin with turns of 1.4°. Skipping arm points that share skin weight with the torso or collar, which protects the armpit's fold, hid real penetration on a physique where the biceps meet the lats: relaxing the skip to 0.40 for the arm points alone takes Gray's audited idle from 100% of frames to none, but the same setting breaks Cadet, whose overlapping armour is built to interpenetrate at the shoulder, and the solver's own measure cannot tell the two apart. So the audit decides, clip by clip. When frames with a part more than 2 cm inside another are still above 3% after the solve, the relaxed solve is run on top, and its turns are kept for each clip in which the audit finds at least two points fewer such frames and no more crushed, stretched or sheared faces (0.05 points) or pops - one more pop where the clip loses twenty points or more, as Gray's landing does (64% to 18%) E156. A first version took or refused the relaxed solve for a whole character, and one clip made worse - Cadet's armour - kept the rest from it; clip by clip Cadet keeps it for his walk, bow and fist pump and goes from 11.5% to 10.3% of frames. The solve poses the arm by linear skinning while the audit and the playground blend dual quaternions; judging the surface the audit's way instead (a differentiable dual quaternion skin, E156) moved the share of deep frames by -0.3 points on Cadet, +0.7 on Pip and +1.2 on Vex, so the solve stays linear. Smooth corrections E165. Measured against the same capture retargeted with neither step, the shipped clips carried 1.3-2.4 times its joint jerk: the solve's turns are found frame by frame and keep frame-to-frame noise, and clearance clamped each key on its own (an elbow gate at 0.05°, a knee floor switching on and off). The corrections, never the capture, are now eased in time: each clip's turns and the shoulder and elbow corrections over 2.5 keys (a zero-phase Gaussian, 42 ms), the knee's over one key and only while the foot is lifted, and the floor bound eased in without being crossed. The added jerk falls to 0.5-1.6% (Cadet 1.28 → 1.005, Mara 2.36 → 1.016, Pip 1.53 → 1.008 times the capture) with the same deep frames, foot slip and pops, at 0.02-0.06 cm more depth at the deepest point. Clips made from video are re-posed by the refine stage after all this, and a punch with the arms crossing the chest goes back into it, so the solve is run again on them and kept for a clip only if its silhouette still matches the video as well (the same IoU refine measures, at the same placement; at most 0.01 lost).

4.11 Moves from generated video

The video. The character's reference image and a written move - timed beats: ready, the move, recovery - go to the community Ref2VA fine-tune [69] based on MiniMax H3 [70], which returns 5.2 s at 24 fps of the character performing it (32-48 minutes on the laptop, E030); its text encoder and VAEs are supplied by a separate Comfy-Org package [71]. Because the performer is the character, its proportions are close to the rig's.

Retrieval. ViTPose [60] reads 17 points in every frame. The sequence is matched against a library of 2,453 captures from the pinned mirror [67] by dynamic time warping [68] of each clip's 2D projection from 16 camera angles, mirrored or not. In a self-test on 40 library clips shown from random angles with pose noise, a stretch and a speed change, the right clip or an exact duplicate ranked first every time, and the angle was off by 5.4° (median) E001.

The fit. A move the library lacks is approximated by its nearest capture, then bent toward the video bone by bone: each bone's direction in the image from the video, its depth from its length - kept from the capture where the two agree within the pose model's noise, and only when the capture misses the video by more than 1.2 times that noise E002. The bone lengths are the video's own, measured where the capture shows each bone flat to the camera E003: with a capture actor's lengths, Aoi's short torso bent into a 28° lean. Table 3 shows the trade-off on the self-test: on the actor's own body the capture's lengths are, as they must be, slightly better; on another body - the real case - the video's own lengths are clearly better. An arm's two depth signs are chosen together, so an elbow cannot bend past what the library's elbows ever do E007.

Onto the character, and refined. The fitted clip is retargeted by copying rotations and then aiming the character's own limbs, spine and head along the fitted bones, which removes rest-pose differences of 4-12° at the thighs and 16-27° at the collarbones E004. Finally the character's own skinned mesh is posed in PyTorch - forward kinematics mirroring Blender's - projected with the video's camera, and 20 bones a frame are optimised against the video in the manner of SMPLify [63], but on the character rather than a body model E005: the silhouette's chamfer both ways, the joints, the face points, DWPose's hands [61] (wrists, forearm twist and 30 finger joints E006), a prior to the fitted clip, smoothness, joint limits, and a bend limit at the elbow. The feet are then put back on the video's floor line E008.

4.12 Moves from words

The capture library has no salute, no boxing guard, no march. A video of the character doing the move works (Section 4.11) but takes half an hour of video generation per move. UniMate [93] generates a skeletal motion from a sentence for a skeleton it has never seen: a flow-matching model conditioned on the skeleton's topology and T-pose as well as on the text, encoded by Flan-T5 [113]. We run it on each character's own skeleton E129. UniMate's own preprocessing reads the packaged character - its canonical T-pose, its topology and its clips as motion features - through Blender 4.0 as a Python module, the version it was written for. Its sampler integrates the flow with an adaptive solver that needs double precision, which Apple's GPU lacks; a fixed 50-step Euler grid runs on the GPU in about 25 s per two-second move. UniMate's exporter writes an FBX on the character's canonical rig, and the move then goes through retargeting, foot planting and clearance (Section 4.9) like any captured clip. Two of our conventions broke it. Face-rig morph targets kept their old coordinates when UniMate moved the mesh into its canonical frame, so the export tore into ribbons even for the character's own clips; UniMate gets a copy without them, since it only makes bone motion. And its pruned skeleton ends at the toe bone, which the FBX importer drops as a leaf, so the generated moves went unplanted; the foot bone's tail now stands in for a missing toe.

Each move is sampled three times and the best kept by eye (Figure 18; automating the choice is E143). Prompts follow the wording of UniMate's training captions ("An object marches in place."): a longer, more specific sentence for the march and for looking around produced lunges and hands on the head, the short one a march.

4.13 Running it on a laptop

Everything runs on a 24 GB M5 MacBook with shared memory. Two jobs on the GPU at once do not merely slow down: two refines corrupted each other's results (Aoi's silhouette 0.672 → 0.647, the kick's 0.733 → 0.485, where each alone reaches 0.82-0.86; one ended in NaN), and a 3D pass beside another job was stopped by the operating system's GPU watchdog. A machine-wide lock now serialises every GPU step of every job E023. The video model is a 33-billion-parameter model; running it here took rewriting its float8 scales for a GPU without float8 and merging its LoRA offline, since patching it at load doubled its memory E030. The Studio (Figure 19) queues jobs, survives its own restart mid-job, and shows each job's stage and time left.

5 Evaluation

5.1 Measures

We measure what a player would see, in units that do not depend on the character's size.

• Joint error (moves from video): the character's clip is rendered at the video's own frame times from the video's camera angle; the same pose model [60] reads the video and the render, the hips are laid on each other frame by frame with one scale for the clip, and the mean distance between the twelve body joints is reported in the video's torso lengths (shoulders to hips; 0.05 is about 2.5 cm on an adult).
• Silhouette IoU: the render's outline against the video's figure, cut out by a salient-object model [98].
• Face error: nose, eyes and ears about their centre, in torso lengths - which way the head turns.
• Hands: DWPose [61] on both, each hand it is sure of; the 21 points about the wrist in palm lengths (mean and median), how often the palm faces the video's way, and the angle between the hands' pointing directions.
• Per stage: the same joint error for the fitted points, the FBX written from them, the character's own skeleton, and the pose model's reading of the render - which separates the rig's error from the scorer's.
• Four views: the clip rendered orthographically from the video's camera, the side and above, at one scale on a 10 cm grid, with each elbow's and knee's bend and each shoe's height above the floor read out per frame.
• Foot slip: while a foot is on the floor, how far its lowest vertex slides between frames, as a share of the gait's actual nonzero ground speed, driven by the playground's own controller; absolute slip in m/s is also retained. A percentage is undefined for a stationary gait. The historical table used a 1 m/s denominator floor and is provisional below that speed; Blender's corrected v2 path and the still-legacy browser audit also need contact-mask parity before the same-schedule rerun E029 E122. Sole depth remains a separate fixed-floor measure, not contact-conditioned penetration. Human perception thresholds from mannequin studies are rating-design evidence, not a universal pass value [77].
• Front silhouette IoU of a 3D model against its reference picture, rendered from the front.
• Aberrations E102: every frame of every clip of every character, on the mesh as the viewers deform it (dual quaternion skinning), against its rest pose. Faces crushed below half their rest area or stretched past twice it; faces sheared, turned more than 60° beyond their own bone; triangles that intersect although they were more than 3 cm apart at rest [114]; penetration, how deep one body region has gone inside another, counted as deep past 2 cm; vertices under the floor; and popping, a vertex whose offset from its neighbours spikes out and back within three frames. Each is assigned to the body region of the bone that dominates the surface there. Shares are of all faces, so characters compare directly.

5.2 Characters

Table 4 lists the 17 characters built so far - 17 in the public playground - with what each was made from and the checks this paper describes. 16 have a face rig; the others' faces could not be read and were refused (Section 4.7). Build times are in Table 1.

5.3 Moves from video

Table 5 follows four moves in three styles from the first measurement to now. The first column is the pipeline as it was before this line of work (the nearest capture, retargeted), the second after fitting on the video's own body and aiming the character's limbs, the third after the refine stage, the hand terms and the floor. Silhouette overlap rose from 0.609-0.751 to 0.771-0.843. At rest the same models lie on their reference pictures at 0.90-0.93, so most of what remains is in the motion.

The hands show why means alone mislead. Aoi's hand error is 0.30 palm lengths in the mean but 0.15 in the median: in about twenty frames her jacket sleeve, which rides the upper arm, sticks out past her bent elbow, and DWPose reads the sleeve's light lining as her hand. That is a skinning limit, not a hand error (Section 8).

Table 6 breaks each move's error down by stage. The fit is close to the video everywhere (0.054-0.119 after the refine, 0.037-0.039 for the fitted points); writing it as an FBX and moving it onto the character adds error, mostly the capture's and the character's body proportions. The last column counts frames where the pose model's reading of the render disagrees with the rig's own skeleton by more than 0.3 torso lengths - the scorer's error, not the rig's (Section 5.12).

5.4 Seeing every move from four sides

One camera cannot see depth. A clip can match its video from the video's angle while an arm folds back through itself behind the figure. We render every move from the video's camera, the side and above at once, beside the video, with the joints read out (Figure 21) E009. It found three faults the scores had missed: Aoi's elbow bent to 176° (the forearm back through the upper arm) for 16 frames, now held at 150° E007; the feet of every move standing off the floor after the refine, Aoi's lowest shoe 3.1 cm up on average, now 0.6 cm E008; and a comparison video running 25% fast E010. It also cleared the rig of a fault the score seemed to show E011.

5.5 Feet

Foot slip is measured on the deformed shoe while the playground's controller carries the character forward at the clip's ground speed: the horizontal speed of the lowest shoe vertex while it touches the floor (within 4 mm), as a share of the gait's true ground speed, and in metres per second (blender/foot_audit.py, metric v2). On 17 characters and nine gaits the worse foot slides a median 3.2% of ground speed, a mean 4.2%, and at worst 28.3% (Table 7). Walking backwards is now the worst gait, at most 10.4% on every character but one. The exception is Cadet: his sprint slides 28.4% and his left strafe 19.1%, the same with or without clearance. His flared greaves come down to within a few centimetres of the floor and are weighted to the ankle, so their rims are planted as if they were the sole: as the ankle flexes a rim holds the floor and the sole floats millimetres above it, and the audit finds his feet down in 8% of sprint keys against a median 24% for the others. Planting only the shoe's contact patch is the next experiment E131.

Which surface of the shoe is planted changes this E130. With a rigid copy of the sole carried on the ankle bone, the 13 characters measured before slid a median 4.01% (mean 4.73%). Posing the whole shoe by its own skin weights but thinning it to 400 points brought the median to 3.64%, and made 12 of the 117 gaits worse by more than a point: walking backwards, Juno 6.0% → 11.4% and Knight2 9.1% → 14.7%; Pip strafing left 15.3% → 20.5%. Their ground speeds were unchanged; the contact slid. Thinned, the lowest point of the shoe jumped from one sample to the next as the foot rolled, and the planted pivot jumped with it. With every vertex the median is 2.89% (mean 3.59%): 39 gaits better than with the rigid sole by more than a point, 2 worse, and the three above at 6.0%, 6.9% and 7.8%. Cadet, whose armoured boots started it, had a boot more than 1 cm under the floor in 48.3% of his frames with the rigid sole, 1.88% thinned, and 1.28% now.

5.6 Aberrations

The aberration audit was checked before it was believed. A synthetic character (a box torso and an arm on two bones with a blended elbow) was built with aberrations of known size, and 22 of 22 checks pass under both skinnings (Table 8): nothing where nothing moves; exactly the share of faces built to be crushed or stretched; the floor at −3 cm; a one-frame pop found on its own frame and a fast smooth bend not called one; and penetrations of 0.5, 1, 2 and 5 cm measured to the millimetre. Three of these checks failed first, and each failure was an error that would have misled the measurements of real characters (Section 5.12). One-frame pops are only seen on the frames sampled, so the audit samples every frame; that costs a fifth more time than every second one.

On the 17 characters, before clearance, 2.72% of faces were crushed, 1.52% stretched and 2.71% sheared, averaged over clips and characters. Crouch walk, the clip furthest from a standing pose, is among the two worst for every character; the other is the sprint or a strafe. Most crushed faces are out of sight: of the flagged faces, a ray along the normal escapes the character for only 26-45% (an armpit collapses under the arm). Dual quaternion skinning, the playground's, crushes fewer faces than linear skinning, the glTF default (2.72% against 3.51%), and stretches more (1.52% against 1.32%): it keeps the volume that linear blending loses, and a joint bulges instead [42]. Penetration was the most common aberration: in 48% of frames, on average, one body region was more than 2 cm inside another (50% counting arms through hair, which the head's region includes).

Holding the clips inside each character's joint limits (Section 4.9) brings that to 29% (30% with hair), and solving every frame against the character's own surface (Section 4.10) then to 5% (6% with hair); the deepest penetration between body regions, averaged over clips, goes from 3.89 to 2.71 and then 1.43 cm (Table 9). Clearance itself raised no character's crushed, stretched or sheared share and added no pop (one hand pop already in Wren's sprint came to straddle two frames) E128, and the contact solve, measured alone, changed those shares by at most 0.04 points and added no pop E151. Against the clips after clearance, this version - the contact solve with its second look and the skin weights chosen at the extreme poses (Section 4.6) - crushes fewer faces on every character (0.92% to 0.74% on average); stretched faces rise on four characters (Juno most, 1.29% to 1.40%), sheared on six by at most 0.05 points, and pops on three (Rivet's 7 to 13, in his crouch walk and jog, not traced yet). The planted feet do not move: Bo's foot slip on walk, jog, run and sprint is identical to the fourth decimal before and after. The largest single turn clearance makes is 7-53° per character, and armour takes the most: Cadet's knee plates meet at 105° of bend, so his sprint's swinging knee, bent 159° in the capture, is opened by up to 53° - a stiffer-legged sprint, plates meeting instead of passing through each other.

What clearance leaves is what a single joint's limit cannot see: hands swinging into thighs that are themselves moving (the limits are measured with the legs at rest), legs brushing past each other in strafes, arms through long hair, which is a job for the hair's own spring chains E125, and knees through a long coat's hem, which needs the hem to follow the thighs.

5.7 Extreme poses

The 18 clips never put the arms overhead or across the chest, or squat deeply, and a neutral stance hides what skinning does there: practitioners' rigging guides ask for checks at extreme poses for that reason, and the open reference-asset-compiler [25] gates its rigging on a five-pose deformation suite E153. blender/deform_suite.py writes ten held poses into a copy of each rig - arms overhead, forward, across the chest and behind; a deep squat; a bend with a twist; a high kick; the head turned hard; and, for the feet, on tiptoe and at heel strike, added when softer weights bled a boot's ankle in a run that the first eight could not see - each turn signed by which way it moves its child joint, so no knowledge of a rig's bone axes is needed, and the audit measures them like clips (Table 10). Skin is mostly sound at the extremes: over 17 characters the median stretches 1.31% of its faces with the arms overhead and crushes 0.43% with the forearms across the chest. The tail is the finding. Vex stretches 4.7% of its faces with the arms overhead, three to four times the median, and the worst for arms forward is Vex (2.86% sheared): the characters in loose jackets and armour, at the torso and upper arm beside the armpit, where a jacket is pulled up with the arm. Penetration at these poses is geometry, not skinning: the forearms across the chest go 7.73 cm into a median torso, and a thicker one takes more. Shares are of all faces, so they compare only between rigs with the same face count: the face rig adds 18,196 faces to Vex's 59,617 and dilutes every share by 23%.

The battery is also what chooses each character's skin weights (Section 4.6): the default falloff and a hybrid, soft on the body and default on the legs, are rigged and posed, and the hybrid is kept when it is at least 0.1 points better on average and no pose is more than 0.3 worse (Table 11). It cannot choose alone. Two weightings it preferred - the learned weights of Make-It-Animatable E157 and distance measured from each part's surface E159 - held Vex's faces in shape at the extremes and in her clips let her arms sink into her jacket two to three times as often, while the second improved every clip measure of Pip's; the battery's own penetration did not tell the two apart, so the clips will have to decide (Section 10).

5.8 Against professionally made models

The measures above are of motion and deformation. Whether a model looks professionally made is measured by a model-quality audit run the same way on every character and on professionally made meshes E132: Blender Studio's Human Base Meshes [115] and Ellie from the open film Sprite Fright [116]. Shape, on the shipped low-poly at the character's height: noise, how far each surface point's normal turns from the mean normal within 1 cm; lumps, how far the surface within 4 cm departs from the quadric that fits it best, the worst quarter left out so a designed crease does not count; triangle tilt, how far each triangle leans against the smooth normals at its corners - what a normal map has to correct; asymmetry, the mirrored surface's distance to the surface about the best mirror plane; open edges and loose pieces. Texture: lighting left in the albedo, the same garment's colour where it faces front and back (CIEDE2000), the colour step across UV seams, and the PBR range. The professional bodies score noise 0.74-0.83°, lumps 0.20-0.27 mm and asymmetry 1.4-1.5 mm, with no open edges and one piece. Ellie, a layered production character, scores noise 2.98° and lumps 1.73 mm: folds and hair are designed detail, and the audit cannot yet tell them from noise, so a garment needs comparing with garments.

Table 12 scores each character's shipped model as version 0.11 built it and as rebuilt on this version's defaults - the finer surface (Section 4.3), the loose shards dropped, the hands and the skin, the albedo range (Section 4.5). Over the 9 characters, the median noise went from 3.07° to 2.92°, lumps from 1.05 to 0.94 mm, triangle tilt from 15.7° to 14.8°, open edges from 11.0 to 5.8 per 10,000 triangles and loose pieces from 13 to 1; the same garment's colour front and back from dE 2.23 to 1.79. The low-poly's numbers move less than the renders do - the solid and the decimation set it - and the gap to the professional meshes remains two to four times on noise and lumps: the visible gain of the finer surface is in the normal map (Figure 5).

Close-ups. A whole model's averages moved little between builds that look very different close up, so every character was rendered in eight lit close-ups - face, three-quarter, a hand from above and both hanging, a shoulder, the upper back, a foot - and each fault listed E138 (Figure 24). On the fourteen version 0.11 builds: faces blotched where the picture is projected onto a face shaped unlike it (nine), shoulders and armpits torn by the T bake (five), two-toned hands (most), rings at bare wrists, a skin-coloured bell at Pip's wrists where the hand was cut beyond his flared cuff, dark streaks in hair, and floating debris (all). Hands, wrists, debris and shoulders are mended above, faces taken part by part below (Section 5.9); hair and Pip's wrists are open (Section 8).

5.9 Faces, on a benchmark set

Faces are judged on a benchmark set rather than the whole roster: two characters for each style and gender - realistic Mara, Juno, Rowan and Knight, stylised Vex, Rivet, Pip and Cadet, anime Aoi, Hana, Kaito and Ren - chosen to differ (hair over the face, a helmet collar, a drawn smile, painted eyes) so a change that helps one kind and hurts another shows E140. Each face is rendered from the front in soft light and scored against its picture two ways. Landmark error: DWPose's 68 face points [61] on the render are fitted onto the picture's by one similarity [106] (the render is framed differently) and the mean distance left, as a share of the eye spacing, is the NME of face alignment. It is not enough: Mara's smeared, squinting version 0.12 face had every point within 3% of the picture's, because a smeared eye is still where an eye should be. Likeness: the render is laid onto the picture by those points and the whole face, the eyes, the nose and the mouth are compared by the cosine of their DINOv3 features [12], robust to the render's light but not to a smeared lid. Over the 10 characters with a version 0.12 build to compare, the median likeness of the eyes went from 0.63 to 0.69 (7 better), of the whole face from 0.67 to 0.71 (9 better), of the mouth from 0.53 to 0.64 (9 better), and the landmark error from 3.2% to 2.7% (Table 13, Figure 25). The gain is the realistic faces': their eyes from 0.51 to 0.63 and whole faces from 0.59 to 0.65, where every part had been misplaced and the eyes were pits. Stylised faces held (0.72 and 0.73) and so did anime (0.82 and 0.83): drawn features were already roughly where the generator sculpts them, one of the eight cannot be read on the model at all (Aoi's eyes are drawn too large; Vex's could not be until her model was made again without the second face on the back of her head, E145), and the faces left wrong are wrong in shape - Ren's head longer than his picture's. On the same characters the rest of the model held: median surface noise 2.62° in version 0.12 and 2.77° now, asymmetry 3.7 and 2.1 mm, faces crushed during the clips 1.15% and 0.97%, foot slip 2.6% and 2.8% of the ground speed; frames with a part deep in another moved by -1.9 to +8.7 points a character (a mean of 2.6): the largest, Vex's, is her arms 2-5 cm into her torso in most clips, at the audit's 2 cm line, and her first rebuild of this round had scored 45.9% against version 0.12's 46.4%.

5.10 Checks a person would make

Two faults shipped that a person saw at a glance and no measure caught: five characters without a right hand, and one with a face on the back of her head. Both are now measured - every modelled fingertip looked for on the low-poly, and a face read on the back of the head E145 E146 - and close-ups of every character's hands, back and back of head are rendered for review. Whether a local vision model can do the looking was measured on the user's own library (LM Studio) E144. Every vision-capable 9B model there read a picture correctly (hair colour, 10 of 10) and told a hand from a stump on 33-34 of 34 close-ups, the five missing hands among them; none picked a move's best sample better than chance. Asked for any defect, the two best found every one labelled by eye - Vex's second face, the pits in three characters' hair, the notches on eight backs - and flagged a third to all of the clean ones too E147. They are used for triage, which close-ups to open first; the measured checks are the gates.

5.11 Moves from words

Cadet's 6 generated moves - the five above and a bow made from the Studio's panel - were audited like his 18 captured ones E129. On the surface they fall inside the captured clips' range: 3.18-4.09% of faces crushed against 1.47-4.96%, 0.61-2.76% stretched against 0.97-4.7%, 1.85-3.74% sheared against 1.83-5.01%, and no pops. The feet were the problem. Before the toe bone was recovered and the shoe planted as the viewer shows it E130, the moves meant to stand still slid: the salute's feet at 0.065 m/s, the ready stance's at 0.579 m/s, against 0.001 m/s for the captured idle. After, 0.002 and 0.088 m/s. The march still slides (0.861 m/s): a high-knee stomp scrapes the toe at every lift and landing, which planting a contact does not remove. Three of the five moves came out as asked on the first wording; the march and looking around needed the short, caption-like one. Generation takes about 25 s per move on the GPU and 11 s to export, against half an hour of video for a move from video; choosing among three samples is still done by eye. Three ways of automating it were measured against the picks made by eye on Cadet's six moves, and none beat chance E143: local vision-language models shown the samples side by side (6-7 of 18 trials, answering by a row's place, not its content); a 9B vision model describing each sample and Jev judging the description (1 of 6); and the skeleton's own motion written up as measured facts for Jev to judge (2 of 6) - exact where vision was not, but the samples differ too subtly for six cases to show it.

5.12 Validity of the measurements

Several of the largest errors we found were in the measuring, not the pipeline. We list them because each would have sent the work in the wrong direction.

• The comparison ran fast E010. The side-by-side and four-view videos resampled the 24 fps video to the renders' 30 fps by taking each video frame once, never twice, so the video played 25% fast and then stood still: the rig seemed to punch 0.2 s late, more as the clip went on. The fidelity score was unaffected, since it renders the rig at the video's own times.
• The scorer misread the render E011. Mara's kick scored 0.071, with a dip to 0.46 at the top of the kick that read as the rig folding its leg. Held against the rig's own skeleton, the pose model's reading of the render was wrong in 7 frames (the skeleton was 0.12-0.14 from the video there, as elsewhere); without them the kick scores 0.049. The four views confirm the leg is at head height.
• Face positions in stairs E019: 3.9 mm quantisation from a half-float render target, found while tracing an unrelated seam.
• Stale caches: the per-stage tool kept its skeleton dumps forever and was scoring a rig two rebuilds old.
• Triangles that change under the audit E102: Blender splits a quad along its shorter diagonal in its current shape, so the synthetic elbow's quads re-split at 55° of bend and a comparison with rest compared different triangles. The audit now splits once, at rest.
• Depth that stopped at 3 cm: the first penetration search looked a fixed six vertex rings from the intersection and measured a 5 cm push as 3.1 cm. It now floods through every vertex that is still inside.
• Inside an open tube: a body region is an open patch (the upper arm is a tube open at both ends), and past its open end the nearest face's normal says nothing about inside. Aoi's bent forearm read 15.9 cm inside her upper arm; Vex's elbow read 15.7 cm at 105° of bend and 0 at 115°. A vertex now counts as inside only if it projects straight onto the face and a ray from it, deeper, leaves the region through its far wall; Vex's elbow then rises smoothly from 0 to 5 cm.
• Intersections are intersections: a review of the audit had renamed its intersecting pairs "broad-phase candidates". A two-triangle test shows Blender's overlap query is a triangle-triangle test after the box test (boxes that overlap without touching are not returned); what it cannot tell apart is contact from penetration, which is why depth is measured.
• Planting that switched itself off E129: foot planting needs the capture's toe bones, UniMate's pruned skeleton ends at them, and the FBX importer drops a final bone as a leaf. Nothing failed: the generated moves were simply never planted, and stood 3-19 cm into the floor, sliding. Retargeting now says so when it cannot plant.
• Planting the wrong surface E130: the planted sole was a rigid copy of the shoe carried on the ankle bone, and the audit measured the mesh. On an ordinary shoe the two agree; Cadet's boot blends with the shin above a high ankle, and the solve kept the copy on the floor while the boot on screen went 4.4 cm under it, in 48% of his frames. Posed by their own skin weights, with linear blending the points still sat a few centimetres above the boot the viewer draws; with dual quaternions, as the viewer skins, they match Blender's deformation within 0.01 mm.

Threats that remain: the joint and hand scores compare two readings by the same pose model, so its blind spots cancel as often as they show; everything is 2D against one camera, with depth checked only by eye in the four views; there are four moves, all from generated video of the character itself, not of people; and times include waits for the GPU. Recovering a known clip from a render of it - a 3D ground truth - is the first experiment of the next phase E101.

6 Negative results

Ideas that looked right and measured worse. Each is kept in the tracker so it is not tried again without new evidence.

7 Comparison with other systems

Table 15 sets CharForge beside the systems and tools it is most often compared with, from each one's own paper or page. It is a comparison of what each does, not yet of how well: measured comparisons on the same characters are experiments E103 (auto-riggers) and E104 (generators).

Three differences matter most. First, where the rig comes from: the research systems either rig a generated mesh with an external auto-rigger (CharacterGen with AccuRIG, DrawingSpinUp with Mixamo's) or avoid rigging a generated mesh at all by fitting a template (Make-A-Character). CharForge rigs the generated solid itself and repairs it for rigging first - arms cut free, hands modelled and sized, garments bound to the bone they move with - because the failures we see are in the mesh an auto-rigger receives, not in the joint placement it would add. Whether a learned rigger such as UniRig [38] would place joints better on our repaired solids is open (E103). Second, fingers and faces: most tools stop at the wrist (Rig Lab treats a hand as one joint) and few rig a face; CharForge models fingers and builds a face rig when it can read the face, and refuses when it cannot. Third, motion: the others retarget library clips; CharForge also makes new ones from generated video and measures them against it.

What we take from them. CharacterGen's multi-view pose canonicalisation addresses the same problem our reference check does (a figure in the wrong pose), upstream; its Poisson blending at view seams is E105. Make-A-Character's library-matched garments and strand hair are one answer to what we cannot do - hair and clothing that move apart from the body E110; its dense landmarks and template face are one answer to faces we refuse E109. The hosted tools' marker-based joint placement is a reminder that a person correcting four markers is cheap; the Studio could offer it.

8 Limitations and open problems

• Fused layers. Hair, braids, bags and layered garments are one surface with what they lie on. Chains on them tear the surface, so Aoi's hair and Mara's braid are rigid, and between Pip's vest and sleeve the armpit stretches with his arms overhead E110.
• A sleeve past a bent elbow. Linear skinning keeps a jacket sleeve that extends past the elbow pointing along the upper arm when the elbow bends far; in Aoi's spell cast the cuffs stick out sideways. Pose-space correctives would reach it E107.
• Joins after the arm cut. 1-3 joined slabs an arm survive the re-cut on several characters; the rig keeps that side of the garment with the arm E026.
• Faces. Faces the stage cannot read get no rig (Bo) E109. Since version 0.13 each part of a face is laid on the model's own and the eyes are rebuilt as lids over a ball E140, but only where DWPose reads the model's face on the generator's paint: most anime eyes are not read and stay as generated, cavities and all, and a face it cannot read at all falls back to the older face flow. The parts are laid on from the front; at the side the face is still the generator's shape, softer than the picture and with no ear or jaw line of its own. The hairline is the hair shell's edge, ragged where the generator left it. A head made again on its own from a close crop of the picture has the shape and remains an option being tested E139.
• The back of the head. A head with a whole second face is caught and made again E145; eye-like pits carved into the back of the hair are not (Ren, Pip, Aoi), and filling them by how enclosed they are turned the hair into blobs. The arm cut's notches on the back are closed to thin lines, not gone E146.
• Close-up faults still open. Hair carries dark streaks where views disagree; Pip's hands are cut beyond his flared cuffs, which leaves a skin-coloured bell at each wrist; armour fused to the arm still stretches when the arm is raised above the shoulder E110 E138. The shipped surface is two to four times noisier and lumpier than a professionally modelled one, and two to three times less symmetric (Table 12).
• Depth in moves from video is borrowed from the nearest capture, so a move unlike anything in the library is right in the image and approximate in depth E108; the pose model loses poses it rarely saw, on video and on render E011.
• Measurement. No independent 3D ground truth for recovered motion yet; current end-to-end motion evidence is four generated clips. The aberration audit finds defects of known size on a synthetic character, but its scores are not yet calibrated against blinded reviewers, so none is used as a gate E101 E102. Its penetration depth assumes each body region is a closed-ish surface: a limb buried whole inside another, with no triangle crossing, is not seen, and a face ray escaping along its normal is a geometric proxy for being in view, not a camera's. Foot slip is now measured against each gait's true ground speed in Blender; the playground's own readout still divides by at least 1 m/s and tracks contact differently, so the two remain different estimands until they are reconciled, and the geometric 4 mm contact test has not been checked against pressure-insole labels E122. Real-video motion and scan-anchored synthetic-view evaluation remain planned E111 E113.
• Generated geometry and texture. Solidification has no source-to-output shape-preservation benchmark E116; retopology has no sampled surface-error score E118; UV checks can miss local stretch despite a good median density E119; and texture gutters have not been checked through the mip levels and samplers used by target engines E120. Part-parser confidence is discarded before label fusion, so its errors cannot yet be calibrated or abstained on E117.
• Prompt-to-asset fidelity. Prompt expansion has rule checks and the model has a front silhouette gate, but the evaluation does not track requested attributes through the reference image into the final character E112. In particular, held props, action/pose requests and numeric height are currently unsupported or omitted by the structured description; short prompts also need a separate passthrough check.
• Independent evaluation. “Held-out” is not yet backed by a frozen subject/sequence/prompt manifest or duplicate-group rule, and repeated frames/views/seeds must not be counted as independent samples. Human quality ratings need blinded, score-boundary examples and reported inter-rater reliability before any proxy is used as a gate E101 E102 E111 E112 E121.
• Engine delivery. The README documents Unity, Unreal and Godot setup, and the paper lists glTF/FBX outputs; version-pinned importer/runtime smoke results are not yet in the experiment record. FBX named LODs and glTF LOD0 have different export shapes, and clip grouping may vary by importer; the manifest does not prove that the optimized web assets or release ZIPs carry every advertised feature. glTF conformance alone will not establish engine behavior E114.
• Speed. A character takes about 41.3 minutes of stage time, a move 35-50 minutes of video generation; most of the former is two 3D passes and two image-model calls E106.
• Licences and data rights. Qwen-Image 2.1 is licensed for non-commercial research/evaluation, with a separate commercial license required [18]. TRELLIS.2's official project page says its materials are for academic/research purposes and not commercial exploitation; this is distinct from any code license [117]. MiniMax H3's Community License excludes the US, EU, UK and Republic of Korea and its use policy restricts outputs outside the applicable territory [118]. The community Ref2VA fine-tune used here is separately labeled Apache-2.0 [69], so compatibility with the upstream H3 terms remains unresolved; neither label alone settles the conflict. This paper is not legal advice. The Mixamo mirror requires acceptance of access conditions, while Adobe's FAQ describes royalty-free use in personal, commercial and non-profit projects; neither source by itself resolves redistribution rights for this mirror [67, 39]. The package can also include original reference images, whose redistribution rights must be checked per character. A complete per-run model/data/version and rights record is still future work; unknown or contradictory terms must remain explicit, not be converted into permission E115.

9 Reproducibility and the living paper

Hardware and software. An M5 MacBook with 24 GB of shared memory; Blender 5.2 [94], ComfyUI [100], Ollama with Gemma 4 [99], TRELLIS.2 through its MLX port [1, 95], PyTorch on Apple's GPU for the pose models and the refine stage. tools/check_env.py reports anything missing across the three Python environments before a long run starts. Model files are not in the repository; the README lists where each comes from and its licence.

Running it. A character: python charforge.py make --prompt "a few words" --name hero, or --image picture.png; or the Studio, python charforge.py studio. A move from video: python charforge.py move --name hero --move punch_combo. The measurements: tools/motion_fidelity.py, tools/motion_stages.py, tools/make_angles.py and blender/foot_audit.py. Built characters are in the playground and as release packages.

How this paper is kept. The paper is built from the repository by paper/build.py. Its prose is in paper/sections/; metric placeholders are expanded from paper/data/metrics.json, which paper/collect.py regenerates from pipeline measurement files, or from paper/data/baselines.json for earlier comparisons. The build checks unresolved placeholders, experiment/citation/cross-reference IDs and missing media. It does not currently detect every manually typed empirical number or prove that prose is linked to a source artifact; some quantitative claims remain hand-entered. Reference metadata and check dates live in paper/references.json, but the build does not fetch and re-verify those source pages. E121 will add a claim-to-artifact ledger and a negative build test; until then, reproducibility is partial, not automated end to end. Experiments are recorded in research/experiments.json before they run - question, method, metric - and completed with their result and decision; the build renders them as Appendix A and research/TRACKER.md. A pipeline change that moves a number should be committed with its data and paper and a line in the changelog (Appendix B). The lab notebook is the day-by-day record the paper is distilled from.

Dependency and data provenance. Model Cards recommend documenting intended uses, evaluation conditions and limitations for released models [119]; Datasheets for Datasets document motivation, composition, collection and recommended uses [120]. W3C PROV-DM [90] and the CycloneDX ML-BOM [91] offer complementary concepts for representing derivations and ML dependencies. CharForge's next step is to carry exact model/checkpoint, input, code/configuration and output lineage into each release, alongside terms and unresolved restrictions; these standards are design references, not a compliance claim. Unknown permissions stay unknown E115.

10 Roadmap

The next phase aims at three things at once: more precise measurement, fewer visible aberrations, and less time per character. The new checks separate known truth, model-based proxies, human acceptance, and engine behavior. Each item is an experiment in the tracker with its metric fixed before it runs.

Where the gaps come from. A self-examination of every measurement over the 17 characters E148 put the largest gaps at the surface (noise and lumps 3.4-4.7 times a professional body's), the motion (27% of animation frames with a body part more than 2 cm inside another - arms and hands in the torso and thighs) and realistic faces (likeness to their picture 0.65), and traced most of the faults found since to one root: the generator makes a single fused surface, and eleven of the pipeline's repairs exist to recover structure it never had, each with failures of its own - missing hands, holes in backs, rings round eyes, lines across mouths. Two root-level changes follow, both from published, openly licensed work. A template body under every character E149: Anny [49], a parametric body with a professional topology, finger and eye joints, UVs and the 52 ARKit facial shapes, fitted under the character in seconds; on Mara its body, wrapped onto her tight clothing, measures surface noise 0.65° and lumps 0.26 mm - in the professional range - and its face, shaped through its own face changes by 68 landmarks, puts her features within 4.5 mm of her model's, where wrapping the face vertex by vertex copied the generated face's faults. What lies outside the body becomes clothing and hair layers: the template fits all six characters tried, stylised and anime builds included (joints within 1 cm, bare skin ~3 mm), takes the character's paint in its own UVs, and with the parser's labels the jacket, trousers, boots and hair come away from the unclothed fit as one layer over the painted body (Figure 27), whose 52 facial actions and painted eyes work on Mara's head (Figure 28). On her rig, with the template's weights folded onto CharForge's bones, it animates worse than the shipped model (crushed faces 0.87% to 1.75%, deep frames 7.5% to 17% on 18 clips) - smooth, tuned weights matter more than topology; rigged by CharForge's own stages instead - garments keeping their own weights, the body hidden wherever the layer covers it, and the contact solve run - and its hand bones taken from the template's own fingers - it has the fewest intersecting surfaces of any version (189 a frame against the shipped 223) and as little stretch (1.04% against 1.03%), with a few more crushed faces (0.90% against 0.86%), pops (5 against 3) and deep frames (5% against 0.25%). On stylised Pip the same chain has fewer deep frames than the shipped Pip (9.0% against 11.2%) and, once its hands are fitted to his modelled knuckles and fingertips, fewer intersecting surfaces (411 against 593), but hands that still deform worse than his modelled ones (pops 31 against 13). It is not in the build yet. Retargeting that sees the body: MeshRet [112], which models the dense interactions between surfaces as it retargets, runs on the laptop but, measured on six characters E150, is no better than clearance and blind to hands in thighs; its idea, solved directly per clip against the character's own surface E151 (Section 4.10), took the roster's frames with a part more than 2 cm inside another from 30% to 8.5% on average (Pip's from 59% to 11%), and now runs in every build.

1. To the commercial bar, in parallel lanes E162: measured against what Meshy and Tripo deliver (Section 2.1), the work is split into lanes that run at the same time, each on its own git branch and worktree, with its own files, cloned test characters and before/after measures, so they cannot overwrite one another (tools/lanes/setup_lane.sh). The lanes:
   • done in 0.19: a sharper texture from the same sources E164, clips that keep the capture's smoothness through our corrections E165, 23 ARKit-named face shapes E166, and cross-faded clips at a fixed simulation rate in the playground E167;
   • dropped: a quad-dominant game mesh. Joining the triangle pairs gives quads that deform slightly worse, and Blender's QuadriFlow refuses our solids E163;
   • dense first, then down to the smallest mesh that keeps the detail - what Tripo does after generating ~1.5-2M faces E168. Measured so far: an isotropic remesh at the same 60k makes far better-shaped triangles but sits 6-8 times further from the surface than our collapse decimation, so the decimation stays. The budget that keeps the surface within 0.35 mm is 40k for Pip, 50k for Mara and Vex and 90k for Cadet's armour; the shipped LODs reuse the full mesh's normal map and their normals are 17-44° off, where LODs decimated from the solid with their own maps stay at 8-17° - next is shipping those, and the face stage's subdivision, which adds half again to every model;
   • the black slots at the back of the shoulders, which the arm cut leaves E169, and a hands check that notices a hand that failed to join the arm;
   • done: Pixal3D against TRELLIS.2 on six characters - an option with a better silhouette and a cleaner surface, not the default for its eyes and symmetry E155;
   • done: a judge that builds a weight candidate through the clips; it rejects the surface-relative weights on all four characters tried, because a thick bone takes a thin joint's weight beside it - a radius per part is next E161;
   • a high-resolution pass on the reference and views, written and waiting on ComfyUI E164.
   Each lane is reviewed by a second agent that re-runs its measures before anything is merged.
2. Shape and texture as clean as a professional's E132: the audit and its baselines now exist, and this version spent the first round on it - the finer surface E135, the loose shards, the hands and the skin E138, the rest pose E137 (Section 5.8). What is left, in the order a viewer sees it: faces - taken part by part in version 0.13 (eyes rebuilt, features laid on the model's own, the face the picture's alone E140); next the hairline, anime eyes the pose model cannot read, the profile, and relief from the picture's estimated normals where it is still missing [121] E136, with a head made on its own as the route to a sculpted shape E139; hair without the dark streaks where views disagree; armour raised above the shoulder, with rigid shoulder pieces E110; the gap of two to four times on surface noise and lumps, which the low-poly's decimation now sets (crease-aware normals, a quad remesh of the solid); and likeness measured beside cleanliness on the whole body as it now is on faces (per-part DINOv3 likeness, Section 5.9), since the texture audit scored Hunyuan3D-Paint's views better on characters where they were plainly worse E134.
3. A 3D ground truth for motion E101: render a character performing a known clip as a video, run the move lane on it with that clip held out of the library, and measure the recovered clip in joint angles, positions, contacts and jitter. Then test independent real footage and subjects E111: TotalCapture and 3DPW supply different video/pose evidence; AMASS supplies varied motion for exact-GT renders, while AIST++ is a reconstructed-motion stress set.
4. Weights chosen by deformation and penetration together [E154, E157, E159]: the extreme-pose battery chooses each character's skin weights by how its faces stretch, shear and crush, and twice preferred weightings that held a loose jacket's faces in shape while the arm sank into it in the clips (Vex: 7% of frames to 22% and 12%). A score that counts both - the battery's penetration beside its deformation, or a few clips audited through the contact solve for each candidate - would let the surface-relative weights take the fins out of raised arms where they do no harm.
5. An aberration score, calibrated against people E102: the audit now measures every frame of every clip and finds known-size defects to the millimetre (Section 5.6); what it lacks is agreement with human judgement. Blinded, randomised ratings of clips near each score's boundary, by at least two raters, before any score becomes a gate; dynamic-human quality studies seed the taxonomy, not a transferable scorer [80, 84, 85]. Texture drift and face-morph deformation need their own measures.
6. What the contact solve does not reach [E128, E151]: the arms are now solved against the torso and legs every frame (Section 4.10); the legs are not - a thigh into a belly in a crouch, shins passing in a strafe - and the moves made from video carry penetrations of 9-11 cm that small turns cannot undo; a bag at the hip that the thigh rises into in a crouch (Wren's satchel carries a third of her frames; re-weighting only moved it from one side of the pair to the other, E158 - it needs a collision response); colliders for the hair's spring chains, so arms push long hair aside instead of through it E125 - and, once hair gets chains, a restoring pull toward the hairstyle's original arc in the head's frame rather than toward the bent parent's direction, as one hair auto-rigger describes [122]; strand-based hair from one picture now exists (DiffLocks [123]: a diffusion transformer decoding one latent code per strand, trained on 40K synthetic styles), but its code needs CUDA kernels, so for now hair stays a rigid shell with spring chains for what hangs; long coat hems that follow the thighs; and corrective shapes at elbows and shoulders, scored by the same audit E107.
7. Moves from words without a person choosing E129: rank UniMate's samples by the audits (E102's aberrations, foot slip) and by how well a render reads as the sentence, so the best of three is picked automatically; write the user's words into the caption style itself; and use UniMate's other two abilities - in-betweening, to join one clip to the next, and editing with chosen joints held, to give a captured walk a new upper-body action without touching its planted feet.
8. Prompt-to-asset fidelity E112: use per-attribute questions, informed by TIFA and GenEval [78, 79], to find where a held-out prompt's attributes are lost or changed between its structured text, reference, and final character.
9. 3D and held-out-view fidelity E113: compare a permission-checked scan set against generated geometry and reserved synthetic orbit renders. Do not call these captured calibrated views; obtain written permission for external sharing of derived THUman data or keep the measurements internal [81, 124].
10. Rigging against the field E103: our joints and weights against UniRig and the hosted auto-riggers on the same meshes, scored by E102; corrective shapes at elbows and shoulders E107.
11. Engine acceptance and provenance E114 E115: run the glTF Validator [86], then smoke-import representative release artifacts in version-pinned Unity, Unreal and Godot projects [87, 88, 89]; report clip coverage, manifest-driven controller error, morph/spring/LOD behavior, runtime costs, exact model/data versions and terms without guessing at permissions.
12. Time E106: each stage timed alone, and the costly ones - the second 3D pass, the detailed face - ablated against their measured benefit; Hunyuan3D 2.1 against TRELLIS.2 on our references E104.
13. Texture seams by Poisson blending E105; depth for moves the library lacks from a world-grounded video model E108; faces the rig will not guess E109; hair and garments as their own layers E110.

The newly identified asset-stage gaps are tracked separately so their measurements can proceed in parallel after the shared data contracts are frozen: solidification topology versus shape fidelity and segmentation confidence E116 E117; retopology distance, local UV distortion and runtime mip bleed E118 E119 E120; 3D hand, expressive-face and independent-layer quality E123 E124 E125. The evidence ledger, corrected foot metric and Studio state-machine acceptance are cross-cutting work E121 E122 E126. The detailed split manifest, artifact paths, formulas and run dependencies remain in the tracker and the development-facing research/GAPS.md. The fast classifier proposed in a user-supplied video is queued as E127 at the very end; its identity and license remain unverified, and it is gated on completion or explicit deferral of the earlier queue.

References

Role: used CharForge runs it · compared weighed or measured against ours · related context. All links checked 26 September 2026; BibTeX.

1. [1] Jianfeng Xiang, Xiaoxue Chen, Sicheng Xu, Ruicheng Wang, Zelong Lv, Yu Deng, Hongyuan Zhu, Yue Dong, Hao Zhao, Nicholas Jing Yuan, Jiaolong Yang. Native and Compact Structured Latents for 3D Generation. CVPR 2026, pp. 14419-14429, 2026. used Published paper for the TRELLIS.2 model used in CharForge: 4B-parameter generation with O-Voxel latents; licensing for the actual model materials is separately recorded at trellis2license.
2. [2] Tencent Hunyuan3D Team. Hunyuan3D 2.1: From Images to High-Fidelity 3D Assets with Production-Ready PBR Material. arXiv:2506.15442, 2025. used Vendored through its Apple Silicon port (hunyuanmac): the shape model makes a character's head on its own from a crop of the reference (E139, opt-in), the paint model the opt-in side and back views (E134); compared as a whole-figure generator in E104. The paper does not establish that unnamed hosted products use it.
3. [3] Hao-Yang Peng, Jia-Peng Zhang, Meng-Hao Guo, Yan-Pei Cao, Shi-Min Hu. CharacterGen: Efficient 3D Character Generation from Single Images with Multi-View Pose Canonicalization. ACM Transactions on Graphics (SIGGRAPH), 2024. compared Four A-pose views from one image, a sparse-view reconstruction model, texture back-projection with depth test and Poisson blending; rigs with AccuRig. Anime3D dataset (13,746 VRoid models).
4. [4] Jianqiang Ren, Chao He, Lin Liu, Jiahao Chen, Yutong Wang, Yafei Song, Jianfang Li, Tangli Xue, Siqi Hu, Tao Chen, Kunkun Zheng, Jianjing Xiang, Liefeng Bo. Make-A-Character: High Quality Text-to-3D Character Generation within Minutes. arXiv:2312.15430, 2023. compared An LLM (Qwen-14B) parses the prompt into attributes; a portrait is generated and fitted onto MetaHuman topology with 431 dense landmarks; strand hair and library-matched garments; 52 blendshapes.
5. [5] Lin Liu, Yutong Wang, Jiahao Chen, Jianfang Li, Tangli Xue, Longlong Li, Jianqiang Ren, Liefeng Bo. Make-A-Character 2: Animatable 3D Character Generation From a Single Image. arXiv:2501.07870 (also Qeios H03RE1), 2025. compared Single portrait to animatable character in under 2 minutes: IC-Light relighting, colour correction, hierarchical face reconstruction, adaptive facial skeleton calibration.
6. [6] Reallusion. AccuRIG: free automatic character rigging. ActorCore, 2025. compared Official material supports marker-guided joint placement and configurable finger counts. CharacterGen reports using AccuRIG; the cited source does not substantiate a generic one-click skinning claim for every standalone workflow.
7. [7] Viggle AI. PINOC AI auto-rigging. viggle.ai, n.d.. compared The product guide describes guide markers and retargeting/mapping for compatible humanoid rigs; its finger controls concern mapping an existing rig, not a claim that PINOC generates finger bones on every output.
8. [8] Annomotion Studio. Rig Lab: AI rigging. annomotion.studio, n.d.. compared Automatic humanoid joints and skinning from an A/T-pose GLB; hands are single joints (no fingers); four procedural clips. The checked page does not establish a release year.
9. [9] Cinevva. Cinevva 3D Model Generator. app.cinevva.com, n.d.. compared The current workbench page advertises text or up to four views, rigging/animation and GLB export with 10K-1.5M triangles; it describes unnamed mesh engines and does not identify Hunyuan3D as its backend. Dynamic, undated vendor page; checked 2026-09-26.
10. [10] Cinevva. Best AI 3D Model Generators in 2026: Text-to-3D and Image-to-3D Compared. app.cinevva.com, 2026. related Vendor editorial guide, published 2026-07-17 and updated 2026-09-14; its topology observations are not a controlled benchmark and it discloses the publisher sells a compared product.
11. [11] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, Jiaolong Yang. Structured 3D Latents for Scalable and Versatile 3D Generation. CVPR, 2025. related TRELLIS, the predecessor of TRELLIS.2.
12. [12] Oriane Siméoni, et al.. DINOv3. arXiv:2508.10104, 2025. used TRELLIS.2's paper identifies DINOv3-L as its image-conditioning encoder; CharForge's local converter references a ViT-L/16 checkpoint, which should be separately pinned in run provenance.
13. [13] Tencent Hunyuan3D Team. Hunyuan3D 2.0: Scaling Diffusion Models for High Resolution Textured 3D Assets Generation. arXiv:2501.12202, 2025. compared An alternative image-to-3D generator (shape DiT + texture painter).
14. [14] AI3DGen. Turn Image to 3D Model Free - No Sign-Up Required. ai3dgen.com, n.d.. compared The vendor page advertises free OBJ exports with selectable 5K-40K face-count choices; it does not substantiate rigging or animation. Free outputs are described as personal-use only, so check terms before project use or redistribution. Page checked 2026-09-26; that is not a release year.
15. [15] Chenfei Wu, Jiahao Li, Jingren Zhou, et al.. Qwen-Image Technical Report. arXiv:2508.02324, 2025. related Technical report for the original Qwen-Image family, not documentation for the Qwen-Image 2.1 checkpoint used by CharForge.
16. [16] Bing Zhao, Chenfei Wu, Deqing Li, et al.. Qwen-Image-2.0 Technical Report. arXiv:2605.10730, 2026. related Technical report for Qwen-Image 2.0 only; it does not document or license the separate 2.1 checkpoint.
17. [17] QwenLM. Qwen-Image-2.1. Official model repository, 2026. used Official repository documents the 7B visual-generation component, native RGBA transparency and ComfyUI support; released 2026-09-20. The exact license is cited separately.
18. [18] QwenLM. Qwen-Image-2.1 Qwen Research License Agreement. Official repository LICENSE, 2026. used The bundled license grants use, modification and distribution only for non-commercial research/evaluation; commercial use requires a separate license. Public source/weights should not be conflated with unrestricted commercial permission.
19. [19] Patrick Pérez, Michel Gangnet, Andrew Blake. Poisson Image Editing. ACM Transactions on Graphics 22(3) (SIGGRAPH), 2003. related Gradient-domain blending, used at texture seams by CharacterGen and Make-A-Character.
20. [20] Yuze He, Yanning Zhou, Wang Zhao, Zhongkai Wu, Kaiwen Xiao, Wei Yang, Yong-Jin Liu, Xiao Han. StdGEN: Semantic-Decomposed 3D Character Generation from Single Images. CVPR 2025, 2025. related Body, clothes and hair as separate meshes from one anime image; Apache-2.0, CUDA. E148.
21. [21] Yuze He, Yanning Zhou, Wang Zhao, Jingwen Ye, Zhongkai Wu, Ran Yi, Yong-Jin Liu. StdGEN++: A Comprehensive System for Semantic-Decomposed 3D Character Generation. arXiv, 2026. related Extends StdGEN: a dual-branch semantic reconstruction model, hybrid implicit surfaces, and video-diffusion texture decomposition into editable layers; body, clothes and hair stay separate. Needs CUDA; the Mac route to the same layering is E149's template body.
22. [22] Weizhe Liu, Yunjie Wu, Xiangqian Shu, Guangwei Wang, Xiangyu Xu, Peng Li, Yujie Li, Hengkai Guo. DreamCharacter-1: From 3D Generative Foundation Models to Product-Ready Character Generation. arXiv, 2026. related Post-trains a 3D foundation model for characters: geometry preference optimisation, two-stage texturing with sparse-voxel inpainting, distillation for speed; reports better ULIP and Uni3D than TRELLIS.2. No code or weights released that we could find.
23. [23] Dong-Yang Li, Wang Zhao, Yuxin Chen, Wenbo Hu, Meng-Hao Guo, Fang-Lue Zhang, Ying Shan, Shi-Min Hu. Pixal3D: Pixel-Aligned 3D Generation from Images. SIGGRAPH 2026 (arXiv), 2026. related Built on the TRELLIS.2 backbone, conditioned by back-projecting image features along camera rays into a 3D volume so the shape is pixel-aligned with the input. MIT; the official code needs CUDA, a third-party C++/GGML port reports about 6 minutes a model on an M5 Mac.
24. [24] grumpyburton. OpenMeshy: local image or text to game-ready 3D. GitHub, 2026. related A peer on Apple Silicon: TRELLIS.2, Hunyuan3D-MLX and Pixal3D backends, retopology, a pixel-match pass that copies the source image's pixels onto the surface, SkinTokens rigging, Unity Humanoid export, and a licence-provenance sidecar with every output.
25. [25] raydeStar. reference-asset-compiler: evidence-gated pipeline from one concept image to a rigged UE5 character. GitHub, 2026. related Gates modelling, retopology, texture, rigging and import on recorded evidence; its rigging gate runs a five-pose deformation suite; no automatic retries; fallback rig routes.
26. [26] Meshy. Meshy API reference: Image to 3D, Remesh and Rigging. docs.meshy.ai, 2026. compared topology quad|triangle (default triangle), target_polycount 100-300,000 (default 30,000; smart topology 100-15,000, default 4,000), texture_resolution 2k|4k|8k (default 2k), enable_pbr, pose_mode a-pose|t-pose, remove_lighting (default true); remesh at /remesh; rigging needs a textured +Z-facing humanoid of at most 300,000 faces and returns rigged FBX/GLB with walking and running clips.
27. [27] Tripo AI, Comfy Org. Tripo image-to-model parameters (Comfy's Tripo node documentation and Tripo developer docs). docs.comfy.org; developers.tripo3d.com, 2026. compared texture_quality standard|detailed|extreme (8K); face_limit -1 adaptive (~1.4M faces v3 standard, 2M detailed); quad output as FBX up to 150,000 faces; smart_low_poly 500-20,000 faces (quad 500-10,000), may fail on complex subjects; auto_size to metres.
28. [28] Jingwei Huang, Yichao Zhou, Leonidas Guibas. ManifoldPlus: A Robust and Scalable Watertight Manifold Surface Generation Method for Triangle Soups. ACM Transactions on Graphics 39(6), 2020. related Repairs triangle soups through voxel occupancy and projection while explicitly balancing watertight topology with resemblance to the input; motivates separate topology-success and shape-preservation measurements in E116.
29. [29] Paolo Cignoni, Claudio Rocchini, Roberto Scopigno. Metro: Measuring Error on Simplified Surfaces. Computer Graphics Forum 17(2), 1998. related Surface-sampling method to compare a source surface with its simplified version; informs E118's area-sampled source-to-retopo distances, not a CharForge-specific acceptance threshold.
30. [30] Dongdong Wang, Boqing Gong, Liqiang Wang. On Calibrating Semantic Segmentation Models: Analyses and an Algorithm. CVPR, 2023. related Studies segmentation confidence calibration, including prediction correctness and domain shift; motivates retaining parser probabilities and calibrating abstention on CharForge-specific annotations rather than assuming softmax confidence is correctness.
31. [31] Pau de Jorge, Riccardo Volpi, Philip H. S. Torr, Grégory Rogez. Reliability in Semantic Segmentation: Are We on the Right Track?. CVPR, 2023. related Evaluates segmentation robustness, calibration, misclassification detection and OOD detection; motivates confidence/abstention checks, but its natural-image datasets do not establish CharForge thresholds.
32. [32] Harrison Jesse Smith, Qingyuan Zheng, Yifei Li, Somya Jain, Jessica K. Hodgins. A Method for Animating Children's Drawings of the Human Figure. ACM Transactions on Graphics 42(3), 2023. related 2D rigging of drawings; the baseline DrawingSpinUp compares against.
33. [33] Jie Zhou, Chufeng Xiao, Miu-Ling Lam, Hongbo Fu. DrawingSpinUp: 3D Animation from Single Character Drawings. SIGGRAPH Asia 2024 Conference Papers, 2024. compared Image-to-3D of a drawing, contour removal and restoration, skeleton-based thinning; rigs with Mixamo's auto-rigger (eight front-view keypoints) and nearest-bone weights, retargets with Rokoko. arXiv:2409.08615.
34. [34] Alican Akman, Yusuf Sahillioğlu, T. Metin Sezgin. Generation of 3D Human Models and Animations Using Simple Sketches. Graphics Interface, 2020. related A VAE from 2D stick figures to 3D point-cloud bodies, and animation by interpolating in its latent space.
35. [35] Cong Wei, Xuanchi Ren, Bryan Chu, Weiming Ren, Huan Ling, Jiahui Huang, Laura Leal-Taixé, Sanja Fidler, Wenhu Chen, Zian Wang, Jay Zhangjie Wu. PixelUMM: Encoder-Free Unified Image and Video Understanding and Generation. arXiv:2609.38597, 2026. compared One decoder-only Transformer (Qwen3-8B backbone, 15.2B parameters with its two experts) that reads and writes raw pixels. Released tasks: text-to-image (256 px), text-to-video (176 x 320, 96 frames), image and video understanding; Linux + CUDA 13 + FlashAttention; weights under the NVIDIA One-Way Noncommercial License. Weighed against every CharForge stage in E142.
36. [36] Ilya Baran, Jovan Popović. Automatic Rigging and Animation of 3D Characters. ACM Transactions on Graphics 26(3) (SIGGRAPH), 2007. related Embeds a template skeleton in a mesh and computes weights by heat diffusion.
37. [37] Zhan Xu, Yang Zhou, Evangelos Kalogerakis, Chris Landreth, Karan Singh. RigNet: Neural Rigging for Articulated Characters. ACM Transactions on Graphics 39(4) (SIGGRAPH), 2020. related Learned joints, bone connectivity and skin weights.
38. [38] Jia-Peng Zhang, Cheng-Feng Pu, Meng-Hao Guo, Yan-Pei Cao, Shi-Min Hu. One Model to Rig Them All: Diverse Skeleton Rigging with UniRig. ACM Transactions on Graphics (SIGGRAPH), 2025. compared Autoregressive skeleton prediction (skeleton tree tokenization) and skin weights; Rig-XL dataset of 14,000+ rigged models.
39. [39] Adobe. Mixamo FAQ. Adobe Help Center, 2021. used Adobe's FAQ (updated 2021-09-14) says Mixamo characters/animations may be used royalty-free in personal, commercial and non-profit projects, including games. It does not identify CharForge's local mirror or by itself resolve redistribution of the mirror/raw clip files.
40. [40] Nadia Magnenat-Thalmann, Richard Laperrière, Daniel Thalmann. Joint-Dependent Local Deformations for Hand Animation and Object Grasping. Graphics Interface, 1988. related Early joint-dependent local hand deformations for animation and grasping; this paper is not the source for the linear-blend skinning formula.
41. [41] Khronos 3D Formats Working Group. glTF 2.0 Specification. Khronos 3D Formats Working Group, 2021. used Defines the glTF package format and its weighted joint-transform linear-blend skinning path; it does not specify FBX behavior.
42. [42] Ladislav Kavan, Steven Collins, Jiří Žára, Carol O'Sullivan. Skinning with Dual Quaternions. ACM SIGGRAPH Symposium on Interactive 3D Graphics and Games (I3D), 2007. used Dual-quaternion skinning reduces some volume-loss/collapse artifacts of linear blending; CharForge's E024 measurement still finds residual collapsed jacket faces.
43. [43] Olivier Dionne, Martin de Lasa. Geodesic Voxel Binding for Production Character Meshes. ACM SIGGRAPH/Eurographics Symposium on Computer Animation (SCA), 2013. related Skin weights from distances through the voxelised body, the idea CharForge's weights follow.
44. [44] Chaoyue Song, Xiu Li, Fan Yang, Zhongcong Xu, Jiacheng Wei, Fayao Liu, Jiashi Feng, Guosheng Lin, Jianfeng Zhang. Puppeteer: Rig and Animate Your 3D Models. NeurIPS 2025, 2025. related Autoregressive skeleton prediction, attention-based skinning with topology-aware joint attention, and a differentiable optimisation that animates the rig from a video.
45. [45] Jia-peng Zhang, Cheng-Feng Pu, Meng-Hao Guo, Yan-Pei Cao, Shi-Min Hu. Skin Tokens: A Learned Compact Representation for Unified Autoregressive Rigging. arXiv, 2026. related UniRig's successor: skinning weights as discrete tokens from an FSQ-CVAE, skeleton and skin as one sequence, refined with reinforcement learning; reports 98-133% better skinning accuracy than prior methods. MIT; the repository needs an NVIDIA GPU with 14 GB, CUDA 12.1 and flash-attn, so it cannot run on this Mac.
46. [46] Zhiyang Guo, Jinxu Xiang, Kai Ma, Wengang Zhou, Houqiang Li, Ran Zhang. Make-It-Animatable: An Efficient Framework for Authoring Animation-Ready 3D Characters. CVPR 2025 (Highlight), 2025. compared A mesh in, a Mixamo skeleton, blend weights and a rest-pose reset out; version 2 builds on the Hunyuan3D-2.1 shape VAE and takes meshes. Code MIT, weights Apache-2.0 per the model card (huggingface.co/jasongzy/Make-It-Animatable), trained on Mixamo characters. Its requirements name CUDA builds of torch-cluster and PyTorch3D; the parts it uses run in plain PyTorch, and CharForge runs it on a CPU (E157).
47. [47] Yi-Hua Huang, Zi-Xin Zou, Yuting He, Chirui Chang, Cheng-Feng Pu, Ziyi Yang, Yuan-Chen Guo, Yan-Pei Cao, Xiaojuan Qi. AniGen: Unified S^3 Fields for Animatable 3D Asset Generation. arXiv:2604.08746, 2026. related Shape, skeleton and skinning generated together from one image as mutually consistent fields. Not run here: a third-party report of a run on an Apple M1 Max rigged a simple mascot and produced no faces for a clothed character with thin limbs.
48. [48] Chao Wang, Guangyao Ma, John Doublestein, Junming Chen, Yiming Lin, Zhaoen Su, Xiaomin Luo, Shiyang Cheng, Jie Shen, Doug Roble, Dilin Wang, Yilei Li, Rakesh Ranjan. OmniFaceRig: Fully Automatic Inner-Mouth-Aware Face Rigging Across Diverse 3D Character Topologies. arXiv, 2026. related Riggability check, multi-model face parsing, dense keypoint-driven template registration, procedurally fitted teeth, gums and tongue with their texture baked, and collision-aware transfer of up to 155 FACS blendshapes; 20-30 s an asset on an A100. The same recipe as E149's template face; its mouth-interior texture is the step E149 lacked. No code located.
49. [49] Romain Brégier, Guénolé Fiche, Laura Bravo-Sánchez, Thomas Lucas, Matthieu Armando, Philippe Weinzaepfel, Grégory Rogez, Fabien Baradel. Human Mesh Modeling for Anny Body. arXiv:2511.03589, 2025. used Parametric body from MakeHuman anthropometry (infants to elders), code Apache-2.0, data CC0; quad topology, 104 bones with fingers and eyes, UVs, 52 ARKit facial actions. Fitted under the characters in E149.
50. [50] Simon Giebenhain, Tobias Kirschstein, Martin Rünz, Lourdes Agapito, Matthias Nießner. Pixel3DMM: Versatile Screen-Space Priors for Single-Image 3D Face Reconstruction. ICLR 2026, 2026. related Two DINO-based networks predict per-pixel normals and UV coordinates that constrain a FLAME fit from one photo. Code released; CUDA 11.8, CC BY-NC 4.0 and FLAME's non-commercial licence - realistic faces only, and not runnable on this Mac.
51. [51] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, Michael J. Black. Expressive Body Capture: 3D Hands, Face, and Body From a Single Image. CVPR, 2019. related SMPL-X jointly models hands, face, and body; the authors emphasize dense features and pose priors, useful context for E006/E109 but not a replacement template for arbitrary stylized meshes.
52. [52] Javier Romero, Dimitrios Tzionas, Michael J. Black. Embodied Hands: Modeling and Capturing Hands and Bodies Together. ACM Transactions on Graphics (SIGGRAPH Asia), 2017. related MANO learns pose-conditioned hand shape corrections from scans; models/data are research-access gated, so use as a methodological reference unless its terms fit.
53. [53] Gyeongsik Moon, Shoou-I Yu, He Wen, Takaaki Shiratori, Kyoung Mu Lee. InterHand2.6M: A Dataset and Baseline for 3D Interacting Hand Pose Estimation from a Single RGB Image. ECCV, 2020. related Real-captured single/interacting hand sequences with 3D pose labels; supports a subject-held-out check that separates image detector accuracy from retargeted hand articulation. Data terms/split must be pinned before use.
54. [54] Darren Cosker, Eva G. Krumhuber, Adrian Hilton. A FACS Valid 3D Dynamic Action Unit Database with Applications to 3D Dynamic Morphable Facial Modeling. ICCV, 2011. related Dynamic 3D facial action sequences with expert peak-frame FACS labels; useful for an expression challenge taxonomy, but human scans and AU labels do not directly validate stylized blendshape appearance.
55. [55] Timo Bolkart, Tianye Li, Michael J. Black. Instant Multi-View Head Capture Through Learnable Registration. CVPR, 2023. related Introduces TEMPEH and FaMoS: the CVPR paper reports 93 subjects and 28 sequences. Data access requires agreement to its license; human face scans are references, not stylized target truth.
56. [56] Hao Li, Ju Dai, Feng Zhou, Kaida Ning, Lei Li, Junjun Pan. AU-Blendshape for Fine-grained Stylized 3D Facial Expression Manipulation. ICCV, 2025. related Introduces AUBlendSet with 32 action units across 500 identities and stylized AU control; informs E124 challenge design, not a validated CharForge benchmark or unrestricted dataset/model license.
57. [57] J. P. Lewis, Matt Cordner, Nickson Fong. Pose Space Deformation: A Unified Approach to Shape Interpolation and Skeleton-Driven Deformation. SIGGRAPH, 2000. related Foundational pose-conditioned corrective deformation; supports E107's elbow/shoulder correctives, while the fitted examples and runtime constraints remain to be measured for CharForge.
58. [58] Qianli Ma, Jinlong Yang, Anurag Ranjan, Sergi Pujades, Gerard Pons-Moll, Siyu Tang, Michael J. Black. Learning to Dress 3D People in Generative Clothing. CVPR, 2020. related Learns pose- and garment-conditioned clothing geometry from scans; relevant to independent garment layers and pose variation, but not a runtime spring solver or stylized-character validation set.
59. [59] Meysam Madadi, Hugo Bertiche, Wafa Bouzouita, Isabelle Guyon, Sergio Escalera. Learning Cloth Dynamics: 3D+Texture Garment Reconstruction Benchmark. NeurIPS Competition and Demonstration Track, PMLR 133, 2021. related CLOTH3D++ pairs animated dressed-human RGB sequences with dense 3D body/cloth and texture annotations; informs challenge design only and remains subject to dataset access terms.
60. [60] Yufei Xu, Jing Zhang, Qiming Zhang, Dacheng Tao. ViTPose: Simple Vision Transformer Baselines for Human Pose Estimation. NeurIPS, 2022. used 17-point body pose on video frames and on renders (checkpoint usyd-community/vitpose-base-simple).
61. [61] Zhendong Yang, Ailing Zeng, Chun Yuan, Yu Li. Effective Whole-body Pose Estimation with Two-stages Distillation. ICCV Workshops, 2023. used 133 whole-body points; CharForge uses its 21-point hands.
62. [62] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, Michael J. Black. SMPL: A Skinned Multi-Person Linear Model. ACM Transactions on Graphics 34(6) (SIGGRAPH Asia), 2015. related The parametric body most human-recovery methods fit; CharForge fits the character's own skeleton instead.
63. [63] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, Michael J. Black. Keep it SMPL: Automatic Estimation of 3D Human Pose and Shape from a Single Image. ECCV, 2016. related Optimisation of a body model against 2D joints with priors - the analysis-by-synthesis idea the refine stage applies to the character's own mesh.
64. [64] Angjoo Kanazawa, Michael J. Black, David W. Jacobs, Jitendra Malik. End-to-end Recovery of Human Shape and Pose. CVPR, 2018. related Regresses a 3D human body mesh with pose and shape from a single RGB image, not just a skeleton pose.
65. [65] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, Jitendra Malik. Humans in 4D: Reconstructing and Tracking Humans with Transformers. ICCV, 2023. related Video human mesh recovery and tracking - an alternative to our retrieval-and-fit.
66. [66] Soyong Shin, Juyong Kim, Eni Halilaj, Michael J. Black. WHAM: Reconstructing World-grounded Humans with Accurate 3D Motion. CVPR, 2024. related World-grounded 3D motion from video, with foot contact - a candidate replacement for depth borrowed from the library.
67. [67] jasongzy. Mixamo animation FBX mirror. Hugging Face Datasets, n.d.. used The dataset card identifies 2,453 animation FBXs downloaded from Mixamo. CharForge's cached snapshot at revision b1c7f4975ea3261d3d0aa2379f6e24754ccde9d8 independently contains 2,453 files in animation/; the mirror is not Adobe-authored and requires accepting conditions including sharing contact information. That count/access page does not establish redistribution permission.
68. [68] Hiroaki Sakoe, Seibi Chiba. Dynamic Programming Algorithm Optimization for Spoken Word Recognition. IEEE Transactions on Acoustics, Speech, and Signal Processing 26(1), 1978. used Dynamic time warping, for matching a video's pose sequence to library clips.
69. [69] WarmBloodAban. Minimax-h3_Singularity Ref2VA fine-tune. Hugging Face model card, 2026. used The model named by pipeline/motion_video.py: a community W4A8 Ref2VA fine-tune, converted/merged locally for inference. Its model card declares Apache-2.0 and identifies MiniMaxAI/MiniMax-H3 as its base. The upstream MiniMax H3 agreement defines H3 Works to include model derivatives and its use policy covers outputs; compatibility/authorization is unresolved, not cleared by the card label.
70. [70] MiniMax. MiniMax H3 model card and Community License. Hugging Face: MiniMaxAI/MiniMax-H3, 2026. used Official source links the Community License. The agreement defines H3 Works to include model derivatives; it excludes the US, EU, UK and Republic of Korea from the Applicable Territory, and its use policy restricts use/reproduction/distribution/display of Outputs outside that territory. It permits contacting MiniMax about separate authorization. The relationship to a fine-tune card declaring Apache-2.0 remains unresolved; this is not legal advice.
71. [71] Comfy Org. Comfy-Org MiniMax-H3 ComfyUI components. Hugging Face model repository, 2026. used CharForge's motion-video client loads the text encoder and VAEs from this packaging repository; it is not the W4A8 Ref2VA transformer checkpoint used for generation.
72. [72] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, Michael J. Black. AMASS: Archive of Motion Capture as Surface Shapes. ICCV, 2019. related Unifies 15 marker-based mocap datasets as rigged SMPL sequences; broad exact-motion source for controlled renders, but it does not provide paired RGB video.
73. [73] Matthew Trumble, Andrew Gilbert, Charles Malleson, Adrian Hilton, John Collomosse. Total Capture: 3D Human Pose Estimation Fusing Video and Inertial Sensors. BMVC, 2017. related Synchronized calibrated video, IMU, and Vicon labels, including about 1.9 million frames; strong primary motion test but a small indoor participant/action cohort.
74. [74] Timo von Marcard, Roberto Henschel, Michael J. Black, Bodo Rosenhahn, Gerard Pons-Moll. Recovering Accurate 3D Human Pose in The Wild Using IMUs and a Moving Camera. ECCV, 2018. related 3DPW supplies in-the-wild footage from a moving phone camera with IMU/video-derived pose labels; use it to test camera-motion generalization and document its annotation/alignment assumptions.
75. [75] Ruilong Li, Shan Yang, David A. Ross, Angjoo Kanazawa. AI Choreographer: Music Conditioned 3D Dance Generation with AIST++. ICCV, 2021. related 1,408 dance sequences, ten genres, nine camera views; 3D motion is reconstructed from multi-view video, so use as an action-diversity stress set rather than Vicon-grade ground truth and exclude the dataset's ignore-list.
76. [76] Lucas Mourot, Ludovic Hoyet, François Le Clerc, Pierre Hellier. UnderPressure: Deep Learning for Foot Contact Detection, Ground Reaction Force Estimation and Footskate Cleanup. Computer Graphics Forum 41(8), Symposium on Computer Animation, 2022. related Uses pressure-insole recordings as foot-contact evidence; supports a separately labeled contact benchmark, not a justification for deriving contact truth from CharForge's geometric floor threshold.
77. [77] Martin Pražák, Ludovic Hoyet, Carol O'Sullivan. Perceptual Evaluation of Footskate Cleanup. Eurographics/ACM SIGGRAPH Symposium on Computer Animation (SCA), 2011. related Controlled perceptual study reports small footsliding detectable in most tested conditions; its mannequin, view and environment setup informs blind-rating stimuli, not a universal CharForge threshold.
78. [78] Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, Noah A. Smith. TIFA: Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering. ICCV, 2023. related Fine-grained prompt-to-image attribute checks via generated questions and visual question answering; a candidate measurement pattern, not a validated scorer for stylized characters.
79. [79] Dhruba Ghosh, Hannaneh Hajishirzi, Ludwig Schmidt. GenEval: An Object-Focused Framework for Evaluating Text-to-Image Alignment. NeurIPS, 2023. related Object-centric checks for co-occurrence, count, position, and color; useful for composing a character-specific prompt-fidelity test.
80. [80] Yujie Zhang, Bingyang Cui, Qi Yang, Zhu Li, Yiling Xu. Benchmarking and Learning Multi-Dimensional Quality Evaluator for Text-to-3D Generation. ICCV, 2025. related MATE-3D: 1,280 textured meshes with human ratings separated into semantic alignment, geometry, texture, and overall quality; its evaluator should not be assumed to transfer to CharForge characters without calibration.
81. [81] THUman project team. THUman 2.0 Dataset. Project dataset and terms of use, 2020. related Pin the exact release: the current repository README also describes THUman2.1. The public release provides scan meshes/textures and rendering recipes, not a set of captured calibrated multi-view RGB views. Terms state non-commercial research use and restrict third-party sharing of scans or derived data without prior written permission; obtain permission or keep derived evaluation internal.
82. [82] Pedro V. Sander, John Snyder, Steven J. Gortler, Hugues Hoppe. Texture Mapping Progressive Meshes. ACM SIGGRAPH, 2001. related Optimizes a shared parameterization across progressive meshes with attention to texture stretch; motivates measuring local UV distortion rather than island overlap/median density alone.
83. [83] Unity Technologies. Mipmaps. Unity 6 documentation, 2026. related Explains per-level texture downsampling and trilinear runtime sampling; motivates E120's distance/mip inspection rather than assuming a base-level gutter prevents runtime bleed.
84. [84] Zicheng Zhang, Yingjie Zhou, Wei Sun, Wei Lu, Xiongkuo Min, Yu Wang, Guangtao Zhai. DDH-QA: A Dynamic Digital Humans Quality Assessment Database. IEEE International Conference on Multimedia and Expo (ICME), 2023. related Human study on 800 dynamic digital humans and ten motion types; explicitly includes binding error and motion unnaturalness, useful as an aberration taxonomy but not a trained scorer validated on stylized generated characters.
85. [85] Yunhao Li, Sijing Wu, Yucheng Zhu, Huiyu Duan, Zicheng Zhang, Wei Sun, Xiongkuo Min, Guangtao Zhai. DHQA-4D: A Large-Scale Dataset and LMM-Based Metric for Dynamic 4D Digital Human Quality Assessment. Pattern Recognition 180, article 114567, 2026. related 32 scanned dynamic humans, 11 distortion types, and human ratings including temporal discontinuity; a recent reference for E102's test taxonomy, not evidence that its evaluator transfers to CharForge.
86. [86] Khronos Group. glTF Validator. GitHub, 2026. related Validates glTF 2.0 structure/resources and emits a JSON report; it does not prove Unity/Unreal/Godot-specific import or runtime behavior.
87. [87] Unity Technologies. Importing Model Files. Unity 6 documentation, 2026. related Documents Unity model, humanoid rig/clip, material, texture, and blend-shape importer behavior; pin the actual Unity version in the smoke fixture.
88. [88] Epic Games. FBX Skeletal Mesh Pipeline. Unreal Engine 5.8 documentation, 2026. related Documents FBX 2020.2, skeletal meshes, morph targets, LODs, and the single-animation-per-skeletal-mesh-file constraint relevant to package layout.
89. [89] Godot Engine contributors. Import Configuration. Godot Engine stable documentation, 2026. related Documents skin influence limits, material handling, animation sampling/grouping, and import-time options; pin the actual Godot release in the smoke fixture.
90. [90] W3C Provenance Working Group. PROV-DM: The PROV Data Model. W3C Recommendation, 2013. related Models provenance as entities, activities and agents with derivation relations; a conceptual structure for per-run lineage, not a compliance claim.
91. [91] OWASP CycloneDX. Machine Learning Bill of Materials (ML-BOM). CycloneDX specification, 2025. related Defines structured inventory/provenance concepts for ML models and datasets; a candidate schema for dependency inventory, not a substitute for per-run artifacts or legal review.
92. [92] Geir Kjetil Sandve, Anton Nekrutenko, James Taylor, Eivind Hovig. Ten Simple Rules for Reproducible Computational Research. PLoS Computational Biology 9(10), e1003285, 2013. related Practical guidance for archiving exact code/data and recording computational procedures; informs claim-to-artifact lineage in E121.
93. [93] Linzhan Mou, Jiahui Lei, Zhiyang Dou, Chenyue Cai, Chaoyue Song, Adam Finkelstein, Szymon Rusinkiewicz. UniMate: One Unified Model to Animate Diverse Skeletons. SIGGRAPH Asia 2026 (arXiv:2609.05415), 2026. used Flow-matching text-to-motion conditioned on a skeleton's topology and T-pose, so one model animates skeletons it never saw; MIT code and weights (github.com/Friedrich-M/UniMate, huggingface.co/Linzhan/UniMate). CharForge runs it on each character's own skeleton for moves the capture library lacks (E129).
94. [94] Blender Foundation. Blender 5.2 command-line manual. Blender 5.2 documentation, 2026. used Version-pinned command-line/background-run documentation; the fact that CharForge scripts each listed stage is evidenced by repository implementation.
95. [95] Noah Lyons. trellis2mlx: TRELLIS.2 on Apple Silicon (MLX). GitHub, 2026. used Upstream MLX port with CharForge's multi-view patch. On 2026-09-26 the local vendor checkout was at cddaf3cb8a9f28956114956ebe754d6661a3f695 with four modified tracked files and one untracked backup; this repository URL alone does not identify the local runtime source.
96. [96] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo. SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers. NeurIPS, 2021. used SegFormer architecture paper; the actual FASHN-B4 human-parsing checkpoint and its NVIDIA Source Code License are documented separately at fashnparser.
97. [97] FASHN AI. FASHN Human Parser. GitHub repository, n.d.. used Repository documents a SegFormer-B4 fine-tuned for 18-class human parsing and says the model inherits NVIDIA's Source Code License for SegFormer; review that license separately from the SegFormer paper citation.
98. [98] Xuebin Qin, Zichen Zhang, Chenyang Huang, Masood Dehghan, Osmar R. Zaiane, Martin Jagersand. U²-Net: Going deeper with nested U-structure for salient object detection. Pattern Recognition 106, article 107404, 2020. used The U²-Net salient-object detector used through rembg for reference/video-frame cut-outs; this citation is not evidence of human-parser accuracy.
99. [99] Ollama. Ollama Gemma 4 E4B model tag. Ollama model registry, 2026. used CharForge's first-choice Ollama tag is gemma4:e4b; pipeline/describe.py permits environment/settings overrides and falls through to later models. The downloaded quantization/digest is not pinned here.
100. [100] Comfy Org. ComfyUI. GitHub repository, n.d.. used Runs image and video workflows; this rolling repository citation is not an immutable revision for the specific CharForge environment.
101. [101] Youyi Zheng, Hongbo Fu, Oscar Kin-Chung Au, Chiew-Lan Tai. Bilateral Normal Filtering for Mesh Denoising. IEEE Transactions on Visualization and Computer Graphics 17(10), pp. 1521-1530, 2011. used The filtering of the generated surface's face normals in blender/mesh_denoise.py (E135).
102. [102] Xianfang Sun, Paul L. Rosin, Ralph Martin, Frank Langbein. Fast and Effective Feature-Preserving Mesh Denoising. IEEE Transactions on Visualization and Computer Graphics 13(5), pp. 925-938, 2007. used The vertex update that moves a mesh to agree with filtered normals, used after bilateral normal filtering (E135).
103. [103] Ken Museth. VDB: High-Resolution Sparse Volumes with Dynamic Topology. ACM Transactions on Graphics 32(3), 2013. used General sparse VDB volume and level-set method. CharForge's use of level sets for solidification, arm cuts and hands is supported by project implementation, not presented as a result of this paper.
104. [104] Michael Garland, Paul S. Heckbert. Surface simplification using quadric error metrics. SIGGRAPH '97, pp. 209-216, 1997. used Blender's collapse decimation, which makes the low-poly; the head's share of its budget is held by a vertex group (E140).
105. [105] Till Kroeger, Radu Timofte, Dengxin Dai, Luc Van Gool. Fast Optical Flow using Dense Inverse Search. ECCV, 2016. used Aligns the reference to renders of the mesh during texture projection (OpenCV's DIS).
106. [106] Shinji Umeyama. Least-squares estimation of transformation parameters between two point patterns. IEEE Transactions on Pattern Analysis and Machine Intelligence 13(4), pp. 376-380, 1991. used The similarity that moves each part of the face onto the model's own (E140) and fits a render's face points to the picture's in the face score.
107. [107] Fred L. Bookstein. Principal warps: thin-plate splines and the decomposition of deformations. IEEE Transactions on Pattern Analysis and Machine Intelligence 11(6), pp. 567-585, 1989. used The smooth field that spreads the face parts' moves over the face (E140).
108. [108] Adobe Substance 3D. The PBR Guide - Part 2: practical guidelines for creating PBR textures. adobe.com/learn, n.d.. used Non-metal base colour no darker than sRGB 30-50 and no brighter than 240, and free of lighting - the range pipeline/albedo_range.py keeps (E132).
109. [109] Vladimir Talyzin. hunyuan3d-2.1-mac-rocm: Hunyuan3D 2.1 without CUDA, on Apple Silicon (MPS) and AMD (ROCm). GitHub, 2026. used Installers and a compatibility layer around Tencent's Hunyuan3D 2.1; its README lists PBR texturing on Apple Silicon as verified at 256 px. Vendored at commit 6f4b63b for the head shape (E139) and the opt-in texture views (E134).
110. [110] three.js authors. three.js r160 (0.160.0). GitHub release, 2023. used The playground's pinned WebGL renderer and animation mixer version (r160/0.160.0); the version-specific release is cited rather than the rolling homepage.
111. [111] Ruben Villegas, Duygu Ceylan, Aaron Hertzmann, Jimei Yang, Jun Saito. Contact-Aware Retargeting of Skinned Motion. IEEE/CVF International Conference on Computer Vision (ICCV), 2021. related Retargets motion to a target character's skinned geometry while keeping self- and ground contacts and reducing interpenetration, with a learned geometry-conditioned network and encoder-space optimisation; CharForge's E128 addresses the same interpenetration with measured per-character joint limits instead of a learned model.
112. [112] Zijie Ye, Jia-Wei Liu, Jia Jia, Shikun Sun, Mike Zheng Shou. Skinned Motion Retargeting with Dense Geometric Interaction Perception. NeurIPS 2024, 2024. compared Retargeting with dense mesh correspondence and an interaction field that prevents self-penetration and keeps contacts; MIT, pretrained on Mixamo characters. E150.
113. [113] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, et al.. Scaling Instruction-Finetuned Language Models. arXiv:2210.11416, 2022. used Flan-T5; UniMate's text encoder is flan-t5-base (Apache-2.0).
114. [114] Blender Foundation. BVHTree Utilities. Blender 5.2 Python API documentation, 2026. related Documents overlap as returning index pairs between trees and ray casts; a BVH overlap result alone is not a documented narrow-phase triangle-intersection verdict.
115. [115] Blender Studio and community contributions. Human Base Meshes v1.4.1 (asset bundle). blender.org demo files, CC0, 2025. compared Professionally modelled base meshes (male and female, realistic and stylised) - the shape baselines of the quality audit (E132). Updated 20 January 2025.
116. [116] Blender Studio. Ellie Pose Library v2.0.0 (asset bundle; the character Ellie from the open film Sprite Fright). blender.org demo files, CC-BY, 2025. compared A production character with layered hair and garments - the quality audit's (E132) clothed professional baseline. Updated 7 August 2025.
117. [117] Microsoft Research and project authors. TRELLIS.2 project materials and research-use notice. Official project page, n.d.. used The official page calls TRELLIS.2 a research project and says its page materials are solely for academic/research purposes and not intended for commercial exploitation. This is distinct from any code repository license; it does not alone resolve generated-output rights. Dynamic page checked 2026-09-26; footer copyright is not treated as a release year.
118. [118] MiniMax. MiniMax H3 Community License Agreement. Official model repository LICENSE, 2026. used The primary license text, dated 2026-08-02, defines the Applicable Territory and H3 Works/Model Derivatives; section V and Exhibit A restrict use of Outputs outside the Applicable Territory. It says users in excluded territories may contact MiniMax regarding separate authorization. Read the full agreement; this citation is not legal advice.
119. [119] Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, Timnit Gebru. Model Cards for Model Reporting. FAT*, 2019. related Recommends documenting intended uses, evaluation procedures, performance characteristics, and limitations of released models; supports the dependency/provenance register, not a license interpretation.
120. [120] Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé III, Kate Crawford. Datasheets for Datasets. Communications of the ACM 64(12), 86-92, 2021. related Proposes documenting dataset motivation, composition, collection, uses, and maintenance; a practical structure for CharForge's motion/scan evaluation provenance.
121. [121] Chongjie Ye, Yushuang Wu, Ziteng Lu, Jiahao Chang, Xiaoyang Guo, Jiaqing Zhou, Hao Zhao, Xiaoguang Han. Hi3DGen: High-fidelity 3D Geometry Generation from Images via Normal Bridging. ICCV 2025 (arXiv:2503.22236), 2025. related Generates geometry through an estimated normal map as an intermediate; the motivation for E136.
122. [122] Cinevva. Hair that keeps its shape (auto-rigging hair with its own joints and secondary motion). Cinevva blog, 2026. related Joints placed along the connected hair surface outward from its root; the restoring pull aims at the hairstyle's original arc in the head's frame rather than at the direction an already-bent parent implies; sphere and capsule colliders. A blog post: no code or evaluation given.
123. [123] Radu Alexandru Rosu, Keyu Wu, Yao Feng, Youyi Zheng, Michael J. Black. DiffLocks: Generating 3D Hair from a Single Image using Diffusion Models. CVPR 2025, 2025. related A diffusion transformer predicts a scalp texture whose every point is a latent code decoded to one 3D strand; trained on 40K synthetic strand hairstyles; Blender and Alembic export. The repository needs CUDA kernels (NATTEN, FlashAttention-2), so it cannot run on this Mac.
124. [124] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, Oliver Wang. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. CVPR, 2018. related Introduces learned deep-feature perceptual comparison against human similarity judgments; use as a secondary held-out-view image metric, not a standalone geometry or animation score.

Appendix A · Experiment tracker

Every experiment, from research/experiments.json. Open a row for its question, method, metric and decision. The same list, with a "resume here" note for anyone picking the work up, is research/TRACKER.md.

Planned and open

Done and dropped

Appendix B · Changelog

Appendix C · How to cite

Cite the version you read; the version and date are at the top of the page and in the changelog.

@misc{charforge2026,
  title  = {CharForge: rigged, animated game characters from a few words or one image, on a laptop},
  author = {CharForge contributors},
  year   = {2026},
  note   = {Living paper, version 0.20 (2026-10-09)},
  url    = {https://majieddd.github.io/charforge/paper/}
}

Appendix D · Archive

Measurements and methods this paper used to report, kept as they were with the reason each was replaced. The lab notebook keeps the longer record, with its own archive at the end.

Foot slip with a 1 m/s floor under the speed E029

The first foot audit divided each gait's contact slip by max(manifest speed, 1 m/s): values at or above 1 m/s were normalised by the stated speed, values below 1 m/s were understated, and percentages for standing still are undefined, so slow entries were left out. On it the worse foot slid a median 3.5% of ground speed, at worst 18.2%, over 78 character-gait pairs. It was replaced by the true-speed audit of Section 5.5, rerun on every character; the old table is frozen in research/data/e029_feet_archived.json. Note that it also predates the rigid sole's replacement E130.

The generated surface at 200,000 faces, unfiltered E135

Until version 0.11 TRELLIS.2 simplified its surface to 200,000 faces, and the solid, the low-poly and the normal map were built on it as it came. Large flat facets on gently curved armour and sawtooth plate edges reached the normal map as a triangle pattern under a grazing light. Replaced by 1M faces and bilateral normal filtering (Section 4.3); the v0.11 scores of every character are the left side of Table 12.

The T baked into the rest pose E137

Until version 0.11 the rig was posed from the generated A-pose to a T, a copy of its armature modifier applied to the mesh, and the posed skeleton made the rest - so that Mixamo's rest and the target's agreed. The linear blending that carries the mesh up to the T cannot be undone by the same blending on the way down: on Cadet it sheared the pauldrons into shards and left slivers in every arms-down frame. Replaced by the T recorded on the armature and used only in the retargeting maths (Section 4.6); --rest T still builds the old way.

Modelled hands shaded from the colour bake E138

The modelled hands took the face's skin tone with ±12% of light and dark from the base colour bake under them - a bake of the generated mesh, where the modelled hand did not exist - which painted pale and dark patches across most hands. Replaced by shading from the ambient occlusion baked on the solid that contains them.

The face laid on by optical flow, or kept from the generator E022 E140

Until version 0.12 the face was laid on through a finer flow of its own, matched on an upscaled crop on lightness for the eyes and brows and on redness for the lips, and refused when it would move features more than 2.5% of the figure's height - on drawn faces a larger warp painted a second pair of eyes. Knight's generated face sat 2.5% from his picture's: refused, the picture went on unaligned, his fringe across one eye and every feature doubled; forced, the eyes smeared into bands. For realistic characters whose face the flow refused, the front view left the face and fringe to the generator's own texture, softer but on its geometry (Figure 29). Accepted, the flow still squinted and smeared Mara's eyes. Replaced by the face laid on in parts (Section 4.5); both remain in the code as the fallback when the model's face cannot be read.

One skin tone over the face E138

In version 0.12 the skin tone evened the face too, to take the picture's shading off a shape not quite the picture's. With the face laid on in parts the shading lies on the right shape, and the tone's colour bands cut the picture's soft cheek shadow into patches; since version 0.13 the face is left as painted.

A rigid sole on the ankle bone

Until version 0.11 the planted foot was a copy of the sole - the shoe's vertices within 3 cm of the floor at rest, at most 160 - carried rigidly on the ankle bone. It is right for a shoe that is rigid below the ankle and wrong for one that blends with the shin, as Section 5.5 measures.