3D and worlds

Modeling, texturing, geometry, reconstruction, materials, scenes and rendering

• articleHow do prompted frontier LLMs compare with fine-tuned small models at next-state prediction in text environments (including TextWorld), and when did predicted rollouts stay consistent with the real environment?
  Corrected 2026-10-24 - see C3-B1131-01; the original statements below are kept for history and are superseded where the correction says so. Li et al. (arXiv 2512.18832 v2) measure exact-match (EM) next-state accuracy in five text environments. Prompted models (zero-shot or 3-shot) reached 0.00-17.70% zero-shot and 11.43-49.12% 3-shot on TextWorld, the narrative-game engine; fine-tuned Qwen2.5-7B and Llama-3.1-8B (trained on 40K GPT-4o trajectories per environment) reached about 70.5%. On ALFWorld/SciWorld fine-tuned models reach 99%+ but zero-shot frontier models 35-65%. Consistency of rollouts (success when replaying world-model actions in the real environment divided by real success) was 0.91-0.96 in the three structured environments for the fine-tuned Qwen world model and 0.67 on WebShop. Fine-tuned models are not "the LLM itself" prompted; the authors say so by contrasting prompting with SFT.
• articleWhat do the Definition magazine interviews with Quite Brilliant and Veles Productions actually claim about time and cost of AI-generated virtual-production environments, and what do they say about limits?
  Corrected 2026-10-24 - see C3-B1127-01; the original statements below are kept for history and are superseded where the correction says so. A May 2026 trade article quotes two named practitioners. Quite Brilliant (London) used AI-generated backgrounds for 15 set-ups over five days and says the result lacks full 3D parallax; its method is a 3D blockout skinned by AI. Veles Productions (Warsaw) says Marble Gaussian splats let a virtual set that took weeks take hours or minutes and calls it "ten times cheaper"; it also lists hallucinations, perspective errors and limited interactivity and says it is phasing Unreal out. These are interview statements; the article gives no measurement method.
• articleWhy does an autoregressive video model get worse the longer it keeps generating, and what are the names and visible symptoms of the failure?
  Corrected 2026-10-24 - see C3-B1094-02; the original statements below are kept for history and are superseded where the correction says so. The primary papers agree on one mechanism: the model is trained on clean or short context but, when it keeps generating, it conditions on its own slightly wrong frames, so small errors compound (called error accumulation or exposure bias). Reported symptoms differ by method: brightness drift (over-exposure or darkening), motion that freezes, colour shift, flicker when a window shifts, and (separately) content that diverges because nothing remembers what left the frame.
• articleWhat evaluation protocol (resolution, clips, frames, FID/FVD scope, quoted versus recomputed baselines) does each later paper state for its nuScenes FID/FVD, so which tables can even be set side by side?
  Corrected 2026-10-24 - see C3-B1051-01; the original statements below are kept for history and are superseded where the correction says so. The stated protocols differ in at least five ways: Drive-WM 150 validation clips x 6 views = 900 segments of 40 frames with the official UCF FVD code; GenAD FID on 6019 frames and FVD on 4369 clips; UniMLVG following MagicDrive with 900 videos of 16 frames; Panacea generating the whole validation set at 256x512 with 8 frames; Delphi single-frame single-view FID and 4/8/40-frame FVD with recomputed baselines marked. DriveDreamer v1 states a 192x128 training size, v2 448x256 for training and evaluation; MagicDrive-V2 lists DriveDreamer at 128x192 with 32 frames. DriveDreamer-2 gives no clip or frame count for FID/FVD in the HTML text searched, and DriveDreamer v2 has no clips/segments sentence at all; DriveDreamer v1 states 150 clips x 8 segments = 1200, but for action-prediction evaluation. DrivingDiffusion says it matched metrics and settings "a...
• articleWhich terms govern the text-safety models bundled with or called by Cosmos-Transfer1: meta-llama LlamaGuard-7b, NVIDIA Aegis Defensive 1.0 and Llama-Guard-3-8B?
  Corrected 2026-10-24 - see C3-B1125-02; the original statements below are kept for history and are superseded where the correction says so. The HF Cosmos-Transfer1-7B snapshot at ef4dc8e3 caches meta-llama/LlamaGuard-7b (Llama 2 Community License per Meta's gating text) and two revisions of NVIDIA's Aegis adapter (card: governed by the Llama 2 Community License). The GitHub code at 5005e823 instead calls meta-llama/Llama-Guard-3-8B (Meta gating text: Llama 3.1 Community License; NVIDIA ATTRIBUTIONS.md embeds the Llama 3 text), which is not in the bundle. The upstream Meta licence and use-policy files were not copied into the bundle snapshot, and the Meta READMEs and licence files are gated; the licence text comes from the public gating-prompt metadata. No legal conclusion is drawn.
• articleHow do engine-exported depth, segmentation and edge passes actually reach these models, and what do the engine and model docs say (or not say) about matching their encodings?
  Corrected 2026-10-24 - see C3-B1071-01; the original statements below are kept for history and are superseded where the correction says so. The one documented engine integration found is CARLA's Unreal Engine 4 branch (merged Aug-Sep 2025): it renders edge, depth and segmentation control videos and sends them with per-control weights to a separately deployed Cosmos-Transfer1 server. Unreal's Movie Render Queue documents EXR passes such as World Depth in world units. No source read states how engine depth units or segmentation colour conventions should be converted to the encoding the Cosmos branches were trained on (depth normalised to [0, 1], randomised-colour masks), so that conversion is an open interface question.
• articleWhat happens inside a streaming video model when the prompt changes mid-rollout, which cache tricks do the papers use, and which failure modes do they report?
  Corrected 2026-10-24 - see C3-B1086-01; the original statements below are kept for history and are superseded where the correction says so. Four papers handle a mid-rollout prompt change by editing the attention cache. LongLive recomputes cached key-value states with the new prompt (KV recache) and reports the trade-off: with no cache the new prompt is followed but transitions are abrupt; with a plain cache visuals are smooth but the new prompt is delayed or ignored. Zing-0.5 refreshes only the text K/V and pins the first frame after the switch; GameCraft-2 uses a prompt recaching mechanism with a sink token; WorldSonus replaces its cross-attention cache at the next 100 ms chunk boundary. MaineCoon gates the next prompt on whether the current spoken line has finished. Reported failure modes: prompt inertia, abrupt transitions, adherence decay across switches, drift beyond about 500 frames and no persistent entity state.
• articleWhat does 'world model' mean in the 2018 Ha & Schmidhuber paper that many later papers cite, and how does its Vision/Memory/Controller design work?
  Corrected 2026-10-24 - see C3-B1125-01; the original statements below are kept for history and are superseded where the correction says so. In Ha & Schmidhuber (arXiv v4) a world model is a learned generative model of an RL environment: a VAE (V) compresses each frame to a latent z, an MDN-RNN (M) predicts the next z given the action, and a tiny linear controller (C) picks actions from z and M's hidden state. The paper shows C can be trained inside M's generated 'dream' and then transferred, and that this fails when C exploits M's errors unless the model's temperature is raised. The meaning is a latent-dynamics model for control on small game environments, not a video or 3D generator. Schmidhuber's 1989-90 'model network' reports are the cited lineage; the phrase 'world model' is the 2018 paper's usage.
• articleWhat do MASS (authoritative learned state) and WorldCast (distributed clients exchanging state) show about moving the shared state out of the video generator, and what do their evaluation protocols leave untested?
  Corrected 2026-10-24 - see C3-B1085-01; the original statements below are kept for history and are superseded where the correction says so. MASS shows, on a matched Snake benchmark at 128x128 and 128 ticks, that a typed learned state recovers more objects (parser 0.764 versus at most 0.128 for video baselines) and has zero cross-view disagreement by construction, but its structural-validity score is worse than independent per-view baselines and its local prediction decays when other players inputs are missing. WorldCast shows on Counter-Strike 2 that exchanging camera-aligned player state raises the rate at which other players are drawn (0.817 versus 0.056 and 0.074 for retrained baselines), but evaluates lockstep clients without network delay or faults and with ground-truth states for most whole-round results. Neither releases code or weights that were found.
• articleWhich open-weight interactive or action-conditioned world models were released after Oasis and DIAMOND (Matrix-Game 1.0-4.0, Hunyuan-GameCraft, LingBot-World 1 and 2), what is actually released for each, and what licence is attached to which part?
  Corrected 2026-10-24 - see C3-B1075-01; the original statements below are kept for history and are superseded where the correction says so. Eight released model lines with public weights were found: Matrix-Game 1.0 (MIT), 2.0 (MIT), 3.0 (Apache-2.0, 5B base and distilled; a 28B model is announced as 'soon'), 4.0 (Apache-2.0; scene checkpoints for three games, released 2026-10-15), Hunyuan-GameCraft-1.0 (Tencent Hunyuan Community License with territory exclusions), LingBot-World 1 (Apache-2.0) and LingBot-World 2.0/Infinity (CC BY-NC-SA 4.0, non-commercial). Licences differ by version inside one project and by component (code, weights, game assets, bundled third-party models). 'Open weights' here means files present on Hugging Face at the pinned revision with a licence text; nothing was downloaded or run.
• articleWhich structural control (depth, segmentation, edges, blur, normals) should I weight to preserve which property, according to the ablations that report it?
  Corrected 2026-10-24 - see C3-B1087-02; the original statements below are kept for history and are superseded where the correction says so. In Cosmos-Transfer1's TransferBench results, blur-visual and edge controls give the highest adherence and lowest diversity, while depth and segmentation allow more variation; raising a modality's weight in a region raised its alignment metric there (Pearson 0.93 and -0.92 in the authors' test). RealMaster's ablation found edges preserved faces better than depth for propagation. EPE found G-buffer inputs matter for realism. Normals appear as a control only in vendor text (Reallusion) and EPE's G-buffers, not in the Cosmos Transfer1 README or VACE's task table.
• articleIn Kang et al.'s controlled 2D testbed, how do in-distribution, out-of-distribution and combinatorial generalisation differ for video diffusion models, and what can not be concluded about commercial video models?
  Corrected 2026-10-24 - see C3-B680-01; the original statements below are kept for history and are superseded where the correction says so. Kang et al. (ICML 2025) train diffusion-transformer video models from scratch on Box2D/PHYRE simulations so the training set is fully controlled. Within the training range the models track the simulator almost exactly; outside it (new speeds or radii) error is about 35x higher and does not shrink with 100x more data or 14x more parameters. Where every object type is seen but not every combination, scaling the number of combination templates (6 to 60) cut human-labelled 'abnormal' out-of-template videos from 67% to 10%. Attribute-pair tests show the models copy the nearest training example, preferring colour over size over velocity over shape. The setting is a 2D toy world, so the paper does not measure Sora-class or other commercial models.
• articleFor each of these open interactive world models, what do the papers and READMEs say about the action space, output resolution, frame rate, context or memory, and hardware, and which cells are measured, claimed or absent?
  Corrected 2026-10-24 - see C3-B981-01; the original statements below are kept for history and are superseded where the correction says so. The numbers are not comparable across systems because they mean different things: Matrix-Game 2.0's 25 FPS is generation speed on one H100 at 352x640; Hunyuan-GameCraft's 25 fps is the video frame rate while its own table lists inference speed of 0.25 FPS (6.6 with distillation) and the text says latency under 5 s per action; LingBot-World 1 claims under 1 s latency at 16 fps; LingBot 2.0 claims 720p at 60 fps without a hardware statement in the text read; Matrix-Game 4.0 takes player input in a game engine and renders low-poly frames, reporting 480p 60 FPS on an RTX 5090 with 112 ms key-press-to-display latency. Memory is explicit only in Matrix-Game 3.0/4.0 (README/report statements) and is stated as absent or emergent in Matrix-Game 2.0 and LingBot-World.
• articleCan an agent run these action-following benchmarks itself: what code, data and checkpoints does each need, in which file formats, and what do the repositories say about terms?
  Corrected 2026-10-24 - see C3-B1046-02; the original statements below are kept for history and are superseded where the correction says so. Per component at the pinned commits: GameWorldScore (code Apache-2.0 text in its directory, MIT at root; needs five external helper checkpoints in .pth/.model/.weights formats); WorldScore (code MIT, dataset on HF with metadata mit and a WorldScore-Dataset.zip in its file list, evaluation checkpoints .pth/.pt/.pkl fetched by wget, leaderboard submission by results or by videos); MineWorld (code MIT, metrics/ folder, .ckpt checkpoints named in the README tree but taken down per the README and the HF API returned 401 on 2026-10-22); CameraCtrl (Apache-2.0 code, .ckpt weights on HF); Yume (Apache-2.0 code, weights on HF per README, processed dataset listed as not yet provided, Yume-Bench data not located); GameCraft (Hunyuan Community License excluding EU, UK, South Korea; .pt weights; no evaluation code in the root listing). No repository text rea...
• articleHow was the multi-player training data in these papers obtained (bots, simulators, game recordings, single-view video), and what does each source imply about generalising to human play and about rights?
  Corrected 2026-10-24 - see C3-B1057-01; the original statements below are kept for history and are superseded where the correction says so. Five data routes appear: scripted Mineflayer bots in Minecraft (Solaris, Gamma-World), publicly available bots in Rocket League (MIRA), a driving simulator (ShareVerse, CARLA), recordings of a co-op game (MultiWorld, It Takes Two, natural play) and of Counter-Strike 2 matches (WorldCast, OpenCS2; whether players were human is not stated in the passages read), and camera-trajectory unrolling of single-view video (MetaWorld). Authors flag bot-only data as a generalisation gap (Solaris, MIRA). Only MIRA states a publisher permission for its gameplay recordings in the text read.
• articleOpen world models for worldbuilding (Oct 2026): persistent 3D worlds (HY-World 2.0) vs memory-based video worlds (WorldCrafter) vs keyboard-controlled video (H3-World)
  Corrected 2026-10-24 - see C3-B1036-01; the original statements below are kept for history and are superseded where the correction says so. Three open releases with different output types. (1) HY-World 2.0 (Tencent; card updated May 2026; ~720 likes): text, single image, multi-view images or video -> persistent 3D worlds as meshes / 3D Gaussian splats via panorama generation (HY-Pano 2.0) -> trajectory planning (WorldNav) -> world expansion (WorldStereo 2.0) -> composition (WorldMirror 2.0 + 3DGS); WorldMirror 2.0 (~1.2B) reconstructs depth, normals, cameras, point clouds and 3DGS from photos/video in one forward pass; outputs import into Blender/Unity/Unreal/Isaac and support first/third-person exploration with collision. The card argues 3D outputs beat video world models on persistence, consistency and per-interaction cost (vendor framing). LICENCE: Tencent HY-World 2.0 Community License does not apply in the EU, UK or South Korea, forbids using the works OR THEIR OUTPUT outside t...
• articleWhat consistency and quality evaluations have the authors of HunyuanWorld 1.0, WonderWorld, WonderJourney and World Labs published, and can their numbers be compared?
  Corrected 2026-10-24 - see C3-B1039-01; the original statements below are kept for history and are superseded where the correction says so. Each source evaluates differently and none measures 3D geometric consistency directly. HunyuanWorld 1.0 reports no-reference image-quality metrics (BRISQUE, NIQE, Q-Align) and CLIP alignment on rendered views against baselines, with no ground truth; WonderWorld's only quantitative table is generation time (9.5 s vs 749.5 s and 798.1 s on an A6000) plus qualitative comparison and a stated failure mode; WonderJourney reports a human preference study (400 participants, binary choices on four axes). The World Labs pages read give examples and claims but no published metric; its TOS disclaims dimensional accuracy. The tables are not comparable across papers (different views, baselines, protocols).
• articleWhich open image-to-3D models can studios use commercially, and are there territorial limits (TRELLIS.2, TripoSR, Hunyuan3D)?
  Corrected 2026-10-24 - see C3-B1038-04; the original statements below are kept for history and are superseded where the correction says so. Checked 2026-10-23: Microsoft TRELLIS.2 (GitHub MIT, pushed 2026-07-10; HF TRELLIS.2-4B card mit) and VAST TripoSR (MIT) are permissive. Tencent Hunyuan3D 2.1 (and 2.0, card "tencent-hunyuan-community") community licence states in capitals that it "DOES NOT APPLY IN THE EUROPEAN UNION, UNITED KINGDOM AND SOUTH KOREA"; the licence grant is "for the Territory only"; use, reproduction, distribution or display of the Works, Output or results outside the Territory is unlicensed; and licensees whose products exceeded 1 million monthly active users must request a licence. The same pattern appears in MiniMax H3 (also excluding the US) (C3-B364-01). ComfyUI ships official tutorials for TRELLIS.2, Hunyuan3D-2/2.1, Pixal3D and TripoSplat (docs index captured in C3-B370), so availability in a tool is not a licence grant. For hosted generators (Meshy, Rodin)...
• articleWhat does MagicArticulate release (skeleton model vs skin weights), and what does its dataset cleaning say about skin-weight data quality?
  Problem. Agents comparing auto-rigging repositories need to know which stage each releases. Conditions. Seed3D/MagicArticulate (CVPR 2025 per README), Apache-2.0 code. Method. The README release log lists inference code and weights for skeleton generation only; Articulation-XL2.0 (about 48K models, split 46.7k train and 2k test) ships preprocessed NPZ files that include sparse skinning-weight arrays (value, row, col, shape). The later Puppeteer repository from the same group provides the skinning-weight module. Data-quality lesson. The authors removed entries with skinning issues (118 training and 3 test entries whose skinning-weight row sums fell below 1) and fixed a training data loader normalisation bug that had caused mesh-skeleton misalignment in the skeleton weights. Failure modes. Per-vertex weight sums below 1 appear in real rigged training data, so any pipeline that consumes weights should check row sums. Limitations. README head only was read; the README may state other li...
• articleCan I use Puppeteer (Seed3D, NeurIPS 2025) to predict skin weights for a mesh and skeleton I already have, what does it need, and what licences apply?
  Problem. Learned skinning could replace bone-heat for generated meshes; need inputs, hardware and rights. Conditions. NVIDIA CUDA setup: README pins Python 3.10, PyTorch 2.1.1, CUDA 11.8, flash-attn and pytorch3d. Method. The skinning module takes .obj meshes plus rig .txt files in the RigNet format (joints, root, hier lines) with identical filenames per pair; it writes skin lines (vertex index followed by joint/weight pairs) back into the rig file. The README says inference needs 4.2 GB of VRAM and that complex topology takes longer. Export to FBX uses bpy 4.2.0. Rights. Repository LICENSE is Apache License 2.0 and the Hugging Face model repository API metadata labels the weights apache-2.0; the Articulation-XL2.0 dataset is labelled cc-by-4.0. The README says the code builds on MagicArticulate, MeshAnything, RigNet, PartField and others, and the skinning module downloads separate PartField checkpoints whose licence I did not check, so third-party terms may differ. Evidence on qual...
• articleDoes Blender's Mirror Vertex Group work on an AI-generated, slightly asymmetric character, and how do I get weights for the opposite bone?
  Problem. Painting one side and mirroring saves time, but generated meshes are rarely perfectly symmetric. Conditions. Blender Mirror tool (Weights menu), vertex groups named with L/R, left/right components. Method. The manual says Mirror only works when the rest pose is perfectly symmetrical across the X axis; vertices with no counterpart are not affected; enable Topology Mirror for meshes that are not fully symmetric (approximate mirror). Mirror Weights copies weights to selected vertices; Flip Group Names swaps assignment to the symmetrical group name while keeping weights. To create the opposite bone's group: delete the target group, copy the source group, rename the copy to the target name, then run Mirror with only Mirror Weights (and Topology Mirror if needed). Failure modes. The tool does not transfer weights to the opposite bone group by itself. Limitations. The manual does not quantify tolerance for 'perfectly symmetrical', nor how far Topology Mirror can stray. Version/dat...
• articleWhich Blender weight-clean-up operators (Limit Total, Clean, Smooth, Transfer Weights) should I run on auto-generated weights, and what are their documented side effects?
  Problem. Auto-weights often leave many tiny influences and noisy transitions. Conditions. Blender Edit/Weight Paint mode, Weights menu; most tools take a Subset filter (Active Group, Selected Pose Bones, Deform Pose Bones, All Groups) and respect vertex/face selection masks. Method. Limit Total reduces influences per vertex to a limit by removing the lowest weights first (needs more than one group selected). Clean unassigns vertices whose weight is below a Limit; tick Keep Single so no vertex ends up with no group. Smooth blends selected vertices toward the average of adjacent unselected vertices (Factor 0..1, Iterations, Expand/Contract) and requires vertex selection. Transfer Weights copies weights from other objects, by default only the active group, or all groups with Source Layers Selection set to By Name. Failure modes. Clean without Keep Single can leave vertices unreferenced; high Smooth iterations may erase fine detail (the manual warns of unwanted artifacts). After Limit T...
• articleHow do I fix weights where many bones overlap, such as the face, using Multi-Paint, and when does it refuse to work?
  Problem. Where more than three bones influence a vertex (mouth, eye loops), painting one bone at a time breaks the balance. Conditions. Blender Weight Paint mode, Armature tab with at least two pose bones selected. Method. Enable Multi-Paint: the brush then edits the combined weight of the selected groups and keeps their ratio. Combine with locking: select a loop of bones with Multi-Paint to edit the falloff between the loop and its neighbours, and lock the neighbours with Lock Relative to edit falloff inside the loop (the manual calls this two independent one-dimensional gradients). With Auto Normalize on, the combined value is the sum of weights; otherwise the average. Failure modes. Multi-Paint cannot paint vertices that have no weight in the selected groups and will not reduce the combined weight to zero; X Mirror is symmetrical only if the weights already were. Workaround from the manual: with vertex selection, run Smooth Weight (Selected Pose Bones, low Factor) to copy a nonze...
• articleDo I have to normalise skin weights in Blender before exporting a rig, and how do Normalize All, Auto Normalize and group locks fit together?
  Problem. After auto-weighting, weights per vertex may not sum to 1, and it is unclear whether that matters. Conditions. Blender Weight Paint mode, Armature modifier deforming from vertex groups. Method. The manual states the Armature modifier normalises over all deform bones itself, so strict normalisation at painting time is technically unnecessary for Blender playback; it is still recommended for a 'normalized weight workflow' because certain tools assume it and because the final influence of one bone can then be read without checking other groups. Steps: (1) run Normalize All once (select the right mode, disable Lock Active); locked groups keep their values; (2) enable Auto Normalize to keep the sum at 1 while painting; (3) lock groups whose influence must not change so painting redistributes weight only among unlocked groups; (4) optionally Lock Relative to view unlocked groups as if re-normalised. Failure modes. With Auto Normalize off, individual weights can be any 0..1 value ...
• articleWhat exactly do Blender's three Armature Deform parenting options do to vertex groups (empty, automatic bone-heat, envelope), and which can silently destroy weights I painted?
  Problem. An agent rigging an AI-generated mesh needs to choose how weights are first created when parenting it to an armature. Conditions. Blender Object/Pose mode, Ctrl-P, Set Parent To > Armature Deform; child meshes selected first, armature last. Method. (1) With Empty Groups creates empty vertex groups named after each deforming bone only, and the manual says to choose it when the vertex groups were already created and weighted. (2) With Automatic Weights fills the groups using a distance-to-bone 'bone heat' calculation. (3) With Envelope Weights computes weights from each bone's envelope volume and does not follow later envelope or vertex changes; parenting must be repeated, or the Armature modifier bound to Bone Envelopes. Groups are only made for bones whose Deform flag is on. Failure modes. The manual says automatic weights can leave parts of the mesh not deforming as expected when overlaps occur for complex armatures; and that same-named vertex groups already on the mesh ar...
• articleWhich Blender Graph Editor tools clean and smooth mocap F-curves (Clean, Decimate, Bake, Gaussian, Butterworth), and what artefacts does the manual warn about?
  Problem. Raw markerless curves are dense, noisy and over-keyed. Conditions. Blender Graph Editor, F-curves of an imported FBX/BVH. Method. Bake Keyframes creates a key at every frame (use before filtering so uneven samples do not skew a filter). Clean Keyframes deletes keys equal to neighbours within a Threshold and the manual says to run it after bulk keying and before hand tweaking. Decimate removes keys that influence the shape least, by ratio or by allowed error. Smooth (Gaussian) has Factor, Sigma and Filter Width. Butterworth Smooth preserves peaks and suits large data but can add a ripple when values change rapidly; lower Frequency Cutoff gives a smoother curve; Samples per Frame resamples first and should be raised to 2 for sub-frame keys such as 60 fps data in a 30 fps scene. Failure modes. Over-smoothing flattens contacts and impacts; ripple around fast changes (manual). Limitations. Manual describes controls, not recommended numbers; apply per channel and keep foot contac...
• articleHow do Niagara Fluids differ from an offline Houdini or Blender sim: what grids and solvers do they use, what is 2D versus 3D versus FLIP, and what does Epic say you should bake?
  Problem: deciding whether a real-time fluid sim in Unreal can stand in for an offline pyro or liquid sim. Conditions: Unreal Engine 5 Niagara Fluids plugin. Method: gases run on grids holding density, temperature and velocity per cell; liquids use a hybrid FLIP (particles plus a grid velocity solve); pressure solve iterations trade accuracy for cost. Epic names simulator types: 2D Gas (much faster, suited to games, lacks complex 3D turbulence), 3D Gas (depth and richer flow at higher memory and GPU cost, for hero effects and cinematics), 2D FLIP (splashes, not pools), Shallow Water (height field for pools and wakes) and 3D FLIP (waves, rivers; computationally expensive). Results can be baked to flipbook textures for cheaper playback, and colliders are static meshes, Geometry Collections or depth maps. Example: a 3D gas torch or explosion baked to a flipbook for a mobile-friendly asset. Failure modes: Epic warns that fluid scenes can be graphically intensive and may crash the GPU on ...
• articleHow does Wave Function Collapse generate a tilemap or bitmap, what does it guarantee, and when does it fail (contradiction)?
  Problem: Generate locally-consistent tilemaps or textures from an example. Conditions: input is a small bitmap (overlapping model, NxN patterns, typically N=3) or a tile list with adjacency rules (simple tiled model). Method: count NxN patterns (optionally with rotations and reflections); start with every output cell in an all-allowed superposition; loop: observe (pick the cell with the lowest non-zero Shannon entropy and collapse it using the pattern frequencies), then propagate the removed options through neighbours; stop when everything is collapsed. Guarantees in the README: (C1) output contains only NxN patterns present in the input; (weak C2) pattern distribution is similar to the input. Failure modes: if propagation zeroes every option in a cell, the run hits a contradiction and returns nothing; the README says deciding whether other valid bitmaps exist is NP-hard, so no fast always-finishing solver exists, yet contradictions are rare in practice (author's statement, not meas...
• articleWhat must retopology preserve if the mesh will be animated, and can automatic remeshers do it?
  Problem: remeshers optimise surface fit, not deformation. Conditions: characters that bend at joints. Method: use quad-dominant topology with edge loops that run around joints and are denser where deformation is greater; Blender's manual says loops are a prerequisite for organic character animation, loops end at poles, and quads deform well, preferred over triangles. Blender states Voxel remesh should not be used to create topology for a mesh that will be deformed, because that topology has to follow the flow of the geometry and no perfect automatic tool exists, so it must be done manually. Failure modes: grid topology ignores loop flow around elbows, knees and eyes; poles near joints pinch. Limitations: the specific 'three loops per joint' rule is forum guidance, not in the official sources I fetched. Version: current manual plus 2.79 page.
• articleWhat does Audio2Face-3D output (mesh vertices or ARKit blendshapes), and how do I get it onto my own character rig?
  Problem. Audio-driven face output must land on a specific character. Conditions. Any character with a face mesh or a blendshape or control rig. Method. The Training Framework docs describe two inference outputs: direct vertex positions on a target facial mesh, or blendshape weights, traditionally on the ARKit standard with custom blendshape sets also supported. The paper adds that raw output is tied to one Audio2Face-3D face identity, so it ships a blendshape solver and an ARKit-compatible model; the 52-name weights can then drive other faces such as MetaHuman or custom rigs. For a rig that is not blendshape based, the docs suggest building blendshapes that correspond to rig controls and mapping weights to controls. The model covers skin, tongue, jaw and eyeballs. Failure modes. A target rig lacking tongue or eye targets gets those channels dropped or must be mapped by hand (inference). Limitations. Mapping quality to non-ARKit rigs is not measured in the sources I read. Version/dat...
• articleMoGe in ComfyUI: what is the licence split and does the MoGe-3 default template have a platform limit the tutorial does not mention?
  Problem. The MoGe tutorial ships MoGe-3 (vitg) as the template default and gives no licence or platform note. Conditions. Mac users or anyone recording licences. Method. Upstream README says MoGe code is MIT except the bundled DINOv2 code (Apache 2.0); the Comfy-Org repack card and the Ruicheng moge-2 and moge-3 vitg Hub cards report MIT. Upstream also says MoGe-3 is not supported on macOS because it depends on FlexGEMM (Triton, no macOS wheels). Failure modes. A Mac run of the MoGe-3 template may fail if the Comfy node uses the same dependency. Limitations. The macOS limit is from the upstream implementation; I did not check whether the ComfyUI node reuses FlexGEMM. Licence statement covers code; the README text does not separately address weights beyond card front matter. Version/date. Fetched 2026-10-24.
• articleWhat licence does Pixal3D carry and does its NOTICE cover the DINOv3 file the ComfyUI tutorial downloads?
  Problem. The Pixal3D tutorial states no licence. Conditions. Using the dino_v3_L_naf clip_vision file with Pixal3D in ComfyUI. Method. Upstream NOTICE says Pixal3D (code, parameters, weights, documentation) is MIT and lists third-party components that keep their own licences: dinov2 (Apache-2.0), TRELLIS.2, Direct3D-S2 and MoGe (MIT). The tutorial downloads a DINOv3 file (dino_v3_L_naf_fp32), which that NOTICE does not name. Also, upstream states default pipeline resolution 1536 (1024 in low-VRAM mode), matching the tutorial's 1536 target. Failure modes. Treating Pixal3D MIT as covering the DINOv3 encoder. Limitations. I did not read the DINOv3 licence, so I state no terms for it; whether the Comfy file is DINOv3-derived beyond its name is unverified. Version/date. Fetched 2026-10-24.
• articleDoes TripoSplat use DINOv2 or DINOv3 in ComfyUI, and which licence applies?
  Problem. The TripoSplat tutorial labels the vision encoder DINOv2 in prose and in the download list, but the file it links is dino_v3_vit_h.safetensors. Conditions. Any agent writing install steps or licence notes for TripoSplat. Method. Trust the file name that matches the upstream card layout (clip_vision/dino_v3_vit_h.safetensors) and treat the DINOv2 wording as a documentation inconsistency. The tutorial says only 'released under an open-source license'; the upstream model card front matter states MIT. Failure modes. Writing a DINOv2 attribution or licence note (DINOv2 and DINOv3 carry different terms) from the tutorial text. Limitations. I did not check which DINO version the graph actually loads, and I did not fetch a DINOv3 licence, so I make no claim about it. Version/date. Fetched 2026-10-24.
• articleDoes the Comfy tutorial's 'metric scale depth in metres' for DA3 Metric-Large mean I can use the output as metres directly? Upstream says the output is canonical metric depth that must be multiplied by focal length.
  Problem. The tutorial describes depth_anything_3_metric_large as giving metric scale depth in metres; the upstream model card qualifies this. Conditions. Using the metric model for fog depth, CG scale or camera solves. Method. Treat the raw network output as canonical depth and apply the focal-length multiplication stated on the model card before reading values as metres; check whether the ComfyUI node already does this before applying it a second time. Failure modes. Double-scaling or skipping scaling gives a wrong absolute scale. Limitations. I did not run the node, so I cannot say whether the Comfy node performs the focal scaling internally. Version/date. Fetched 2026-10-24.
• articleWhy do AI-generated meshes often fail print checks, and which Toolbox check maps to each failure?
  Problem. Text-to-3D meshes often arrive unprintable. Conditions. Marching-cubes-like or fused-surface outputs; Method. Inference only: the failure classes map to named checks: holes and flipped faces = Solid; overlapping shells = Intersections; sliver triangles = Degenerate; paper-thin cloth, hair or fingers = Thickness; floating details = Overhang. Run Check All, then Make Manifold, then re-check and scale. Failure modes. Auto-repair on large openings can invent geometry. Limitations. Reasoning not a measured failure rate; no AI mesh was tested; sources document only the checks, not AI origin. Version and date. Blender 3.6 manual, 2025.
• articleDoes TRELLIS's licence say anything about selling physical products made from its outputs?
  Problem. Same merchandise question for TRELLIS. Conditions. microsoft/TRELLIS LICENSE file on main. Method. The file is the MIT licence for the Software: permission to use, copy, modify, publish, distribute, sublicense and sell copies of the Software, with notice retained and no warranty. It has no output clause. Failure modes. MIT on code does not resolve training-data restrictions on weights or datasets (see C3-B088-W18-03 and C3-B149-G53-03 which cover that). Limitations. I did not re-read the README or model cards. Not legal advice; summary of licence text only. Version and date. main at fetch.
• articleIf I 3D-print and sell a figurine made with Hunyuan3D 2.1, what does its licence say about outputs, territory and user numbers?
  Problem. Physical merchandise from AI meshes. Conditions. Hunyuan 3D 2.1 Community License (release June 13, 2025). Method. Read: (a) Tencent claims no rights in Outputs, users are responsible for them; (b) licence does not apply in the EU, UK and South Korea, and use or distribution of Output outside the Territory is unauthorized; (c) above 1 million monthly active users on the release date a separate licence is needed; (d) Output must not be used to improve other AI models. Failure modes. Selling prints into excluded territories or relying on EU-based use can fall outside the licence. Limitations. Text says nothing specific about physical goods; whether the printed product counts as Output 'distribution' is an interpretation (inference). Copyright in AI outputs and training-data issues are separate. Not legal advice; summary of licence text only. Version and date. 3D 2.1 licence file at main.
• articleHow do I bake mesh maps in Adobe Substance 3D Painter before layering paint over an AI-textured mesh, and what cage problems cause artifacts?
  Problem. Smart materials and masks need baked maps (normal, AO, curvature); AI meshes lack a clean high/low pair. Conditions. Substance 3D Painter, mesh with UVs. Method. Switch to Baking mode (F8), tick texture sets, pick bakers, set resolution, dilation and high-poly meshes; choose cage method (distance based, automatic experimental, custom file); press Bake; read the Baking Log. If no high-poly mesh, tick Use Low Poly Mesh as High Poly Mesh. Failure modes. An intersecting cage gives artifacts. Limitations. The helpx.adobe.com baking page returned HTTP 404 to plain curl (so not read); the experienceleague mirror was used. Layer and fill-layer workflow not read from official docs (only search-snippet secondary hints, not quoted). Version and date. Last update October 8, 2026.
• articleCan I paint over an AI-generated base texture in Blender, and what UV limits break it?
  Problem. AI texturing (Paint3D, Hunyuan3D-Paint) leaves seams, baked shadows or wrong detail to fix by hand. Conditions. Mesh with a UV map and the AI texture loaded as an Image Texture node. Method. Texture Paint mode paints into the UV-mapped image; save the image (or pack it) or round-trip via an external editor (GIMP, Krita): edit, re-save, then reload in the Image Editor. Failure modes. Overlapping UVs are not generally supported, which matters when AI UV atlases reuse islands: a stroke across faces sharing texels paints both. Limitations. Texture slot, brush and projection details not read in depth. Inference: re-bake or re-unwrap overlapping islands first. Version and date. Blender 5.2 LTS manual (updated 2026-10-23).
• articleWhat licences cover MeshLab, PyMeshLab and the Blender manual text, and what follows for a studio that scripts repair?
  Problem. Choosing repair tools for a pipeline needs licence clarity. Conditions. Using the tools internally or bundling them. Method. The LICENSE files in the MeshLab and PyMeshLab repos are GNU GPL version 3, a copyleft licence: distributing a product that bundles or links them carries source-offer duties (inference, general GPL practice; confirm with counsel). Using the tool to process your own meshes is generally not a distribution. The Blender manual pages state CC-BY-SA 4.0 for the documentation text (not the software licence). Failure modes. Assuming a tool's licence governs the mesh it outputs: the GPL text read here does not say that (inference). Limitations. File named LICENSE does not itself name the project; I did not check for dual-licensing or per-file exceptions, nor the 3D Print Toolbox add-on's own licence. Not legal advice; summary of licence text only. Version and date. main branches at fetch.
• articleWhich PyMeshLab/MeshLab filters repair non-manifold geometry and duplicates, and which only select problem faces?
  Problem. Scripted repair of a batch of generated meshes. Conditions. PyMeshLab (Python bindings of MeshLab filters). Method. Detect: compute_selection_by_self_intersections_per_face (selects self-intersecting faces), compute_selection_by_non_manifold_edges_per_face. Repair: meshing_repair_non_manifold_edges (removes faces or splits vertices, so each non-manifold edge chain can become a border), meshing_repair_non_manifold_vertices (splits vertices until 2-manifold), meshing_remove_duplicate_vertices (merges same-coordinate vertices), meshing_close_holes, meshing_re_orient_faces_coherently. Failure modes. Removing smallest-area faces opens holes: follow with close-holes then re-check; self-intersections are only selected, not removed (inference: need remeshing or manual cleanup). Limitations. Parameters not exercised; no AI-mesh test run. Version and date. readthedocs latest at fetch.
• articlePrusaSlicer says it auto-repaired errors: how far can I trust slicer auto-repair on a broken mesh, and what external fixes does Prusa name?
  Problem. A slicer imports a non-manifold STL and reports repair. Conditions. PrusaSlicer or Slic3r. Method. Prusa's KB says automatic repair works to a limited degree and a dedicated service may be needed; it names a Windows 10 'Fix by the Netfabb' route plus free external tools (Aspose, Formware, Meshmixer, 3D Builder). It rates Meshmixer Make solid as almost always working but with resolution loss. The older Slic3r manual says some problems are out of the slicer's reach. Failure modes. Third-party tool availability changes (inference: the Netfabb route may be dead: a reader comment on the page claims so, unverified). Limitations. The KB page is undated; Slic3r manual is predecessor project. Version and date. Pages read at fetch time.
• articleHow thin can a wall be for an FDM print, and does a slicer's thin-wall option rescue AI-mesh details thinner than one perimeter?
  Problem. Generated figures often have fine detail and paper-thin parts. Conditions. FDM with a 0.4 mm nozzle (PrusaSlicer, ~0.45 mm extrusion width). Method. Prusa's table gives approx wall thickness per perimeter count: 1 = 0.45 mm, 2 = 0.9 mm, 3 = 1.35 mm, 4 = 1.8 mm (settings-dependent). The Detect thin walls option merges two paths into one pass. Failure modes. Walls below one perimeter are not printable and are dropped or mis-printed; thin-wall detection does not fix that. Limitations. Values assume the stock nozzle; resin printers have other limits not covered. Inference: scale a whole model down and many features cross this limit, so check thickness after resizing. Version and date. Page undated; MK3S-era numbers.
• articleWhat does the Toolbox's Make Manifold clean-up do, and what does the Blender extension listing say it is for?
  Problem. After checks flag holes or flipped faces, one needs a first-pass fix. Conditions. Same add-on. Method. Clean Up > Make Manifold attempts to fix bad normals, fill holes and remove empty edges and faces; Clean Up > Distorted triangulates distorted faces; the manual suggests Merge > By Distance for degenerate geometry and a Boolean modifier for intersecting cubes. Scale To can fit a model to an exact volume or bounding axis. Failure modes. The manual says 'attempts', so complex AI topology (many holes, internal shells) may stay non-manifold: rerun Solid after the fix. Limitations. No success rate is given; inference: automatic hole-filling can create wrong geometry on big openings, so inspect. Version and date. Blender 3.6 manual (2025-06-23) and extension listing.
• articleWhat does Blender's 3D Print Toolbox check on a mesh before slicing (Solid, Intersections, Degenerate, Distorted, Thickness, Edge Sharp, Overhang), and what does each flag mean?
  Problem. A generated mesh looks fine on screen but a slicer needs a closed (watertight) surface. Conditions. Mesh selected in Blender with the 3D Print Toolbox enabled (bundled add-on in 3.6; listed as a Blender extension later). Method. Open 3D Viewport > Sidebar > 3D-Print, run Check All or single checks; in Edit Mode, clicking a result selects the offending parts. Solid flags non-manifold edges (edges not shared by exactly 2 faces) and bad contiguous edges (a flipped face). Intersections, Degenerate (zero area/length), Distorted (non-flat quads/ngons that triangulate unexpectedly), Thickness (thin faces), Edge Sharp and Overhang are separate checks. Failure modes. Per the manual, some slicers tolerate intersections, so a flag is not always fatal; results are analysis only, not a repair. Limitations. English page read is the 3.6 manual (latest path 404 at fetch); no numeric thresholds were on the page text read; check defaults live in add-on settings not read here. Version and dat...
• articleHow does WorldRoamBench test whether an interactive world model remembers a place after revisit, why does it avoid frame-pair matching, and what did it find about memory decay?
  Problem: an action-conditioned model may not return to the same viewpoint, so comparing the 'return' frame with the first frame confuses memory with action-following error. Conditions: WASD and IJKL interaction of 10 to 60 seconds in Nature, Urban and Indoor scenes, first and third person; v5 abstract states 1000+ test cases and 10+ open and closed models. Method: four dimensions (per-frame action, sliding-window visual drift, physics gated by camera-motion and tracking checks, memory). Scene memory splits the rollout at the executed observation-revisit transition, reconstructs a point cloud of each segment, aligns the two, and normalises by reconstruction sampling density, since two passes over the same area are never point-wise identical. Subject memory uses tracking plus a VLM. Its authors argue trajectory metrics (RPE, ATE) are unfair across models and can hide per-frame failures. Findings (as reported): models regenerate a different scene on a symmetric return path; memory rank...
• articleWhat does Reallusion say AccuRIG can rig, in what formats, and what does the free version leave out?
  Problem. Choosing an auto-rigger for a stylised or non-standard character. Conditions. Vendor claims, not independent tests. Method. The AccuRIG page claims it handles oversized heads, obstructed shoulders, atypical limbs and held accessories, can generate hind legs for creature characters, adds extra bone segments to limit forearm twist artefacts, and saves FBX and USD. Its free-versus-advanced comparison table marks one-mesh A/T pose as supported, shows 'allow to import but cannot rig separately' for multi-mesh characters in the free version, and shows Generate Bones as not available in the free column. Use. Treat as a shortlist input; verify on your own mesh using Blender weight-QA steps (C3-B572-08) and test hands, fingers and props. Failure modes. Accepting marketing claims; assuming A/T-pose requirements are optional. Limitations. Marketing page: no limits section, no test data, no version stated; independent docs (Reallusion manual) were not retrievable. Mixamo auto-rigger he...
• articleWhat are the steps and prerequisites for generating a Rigify rig in Blender, and what does Rigify not do?
  Problem. Rigging controls (FK/IK, spine, fingers) are repetitive. Conditions. Blender with the Rigify add-on enabled. Method. (1) Add a predefined meta-rig (Basic Human without face and fingers, Basic Quadruped, Human, Cat, Wolf, Horse, Shark). (2) Fit the meta-rig bones to the character, in Edit Mode or Pose Mode (then Apply Pose as Rest Pose). (3) Click Generate Rig in the armature properties. Preconditions from the manual. Rigify assumes 1 unit is 1 metre (a human is about 2 units tall in Blender's units as the manual puts it); keep legs and spine straight in front view; give knees and elbows a slight bend so Rigify knows their direction; remove face bones if you animate the face with shape keys. Bone rolls should be consistent (X axis is the bend axis). Strengths. A generated rig no longer needs the add-on, and you can re-generate without losing features or animation data. Failure modes. Mis-set bone rolls or collinear limbs. Limitations. Rigify only automates controls and bones...
• articleCan I relight a Gaussian splat scene, and what does the PlayCanvas proxy-mesh method need and not handle?
  Problem. Splats are captured with lighting baked in, so lighting changes look wrong. Conditions. Real-time (PlayCanvas engine) viewing; not a general offline relight. Method. (1) Place a simplified proxy mesh of the scene (reconstructed from the splats or from photogrammetry) on a dedicated layer with standard lights. (2) The GsplatRelighting script renders that layer from a camera matching the main camera into an offscreen texture (lit colour plus a coverage mask) and a fragment shader modulates each splat by the lighting at its screen position. (3) Tune blend (0 original, 1 fully modulated), brightness (default compensates 0.5 grey albedo), and a background multiplier for uncovered splats such as sky. Failure modes. Where the proxy mesh silhouette diverges from the splat surface, lighting boundaries land in the wrong place; uncovered splats are left untinted unless scaled. Limitations. This modulates baked colour rather than recovering material or albedo; hence the baked shadows a...
• articleHow do I clean floaters and background noise from a trained splat in SuperSplat without destroying the original, and what does editing leave out?
  Problem. Raw splats contain floaters and stray Gaussians. Conditions. Browser-based SuperSplat Editor (open source) on a trained splat. Method. (1) Select the splat in the Scene Manager first: selection tools affect only the active splat. (2) Choose a tool by task: Flood selection for isolated floaters, sphere or box volumes for 3D regions, Brush or Lasso for organic areas. (3) Turn on Selection Depth to select only the visible surface when cleaning without punching through; Selection Footprint selects by projected ellipse so large soft Gaussians at an edge are included. (4) Lock finished areas so they cannot be selected or deleted. (5) Deletion is reversible during editing via Undo or Select > Reset, but removing a Scene Manager row is not undoable. (6) Fully transparent Gaussians can be selected, so invisible leftovers that cost memory can be removed. (7) PlayCanvas separately lists floaters, background noise, oversized files and artifacts as reasons to edit, and offers a splat-tr...

More