3D and worlds

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

• articleFor each 4D generator, which components (code, weights, datasets, third-party priors, outputs) have stated terms, what do those texts actually address, and where do texts disagree?
  Rights differ by component and no system is cleared by one licence. Code: 4D-fy Apache-2.0, L4GM Apache-2.0 with embedded third-party notices, Consistent4D Apache-2.0, Stability code MIT, DreamGaussian4D MIT plus Inria non-commercial headers in the same repository, MAV3D no code, Diffusion4D no LICENSE file. Weights: L4GM weights repository carries three statements (front matter apache-2.0, body and LICENSE file CC BY-NC-SA 4.0); SV4D 2.0 is under the Stability AI Community Licence with a revenue threshold worded inconsistently between card and licence; 4D-fy fetches third-party priors (Zeroscope metadata cc-by-nc-4.0; MVDream metadata openrail). Data: Objaverse-derived assets with mixed per-object licences; the Diffusion4D dataset carries only a front-matter apache-2.0. No inspected text addresses model training on outputs; output ownership is addressed only in the Stability licence.
• articleWhich multi-agent or multiplayer generative world models can I actually download or run, and what do the repositories and model cards say about their licences?
  Of eight multi-agent world-model papers inspected, Solaris (two-player Minecraft) ships code, evaluation data, training data and weights under Apache-2.0 per its repository and card. Gamma-World ships Apache-2.0 code and .safetensors checkpoints, but its model card states no licence. MultiWorld states code Apache-2.0 and data/weights CC BY-NC 4.0. MIRA releases code and a gated CC BY-NC-SA dataset but advertises no weights in the text read. MetaWorld releases only single-person code. No release was found for MASS, WorldCast or ShareVerse.
• articleI have engine or simulator renders and want photoreal video through NVIDIA Cosmos-Transfer: what do I feed it, what comes out, and what do the control weights do?
  Cosmos-Transfer1-7B (paper, Mar-Apr 2025) and Cosmos-Transfer2.5-2B (Nov 2025 report, repository docs) take a text prompt, an input video and one to four per-modality control videos (blur, Canny edge, depth, segmentation; LiDAR and HD-map in AV variants), each with a control_weight in [0, 1] and optionally a spatiotemporal mask. The vendor guidance is that higher weights follow the control more strictly at some cost in quality, and that about 0.5 is a middle starting point. Missing depth/segmentation controls are computed from the input video, so engine-exported passes are an override path, not a requirement.
• articleWhat is VideoPhy-2, how are videos scored, how do systems score, and how far can its automatic evaluator be trusted?
  VideoPhy-2 (arXiv v1, Mar 2025) tests physical commonsense on 197 real-world actions with 3,940 LLM-written prompts. Three human annotators rate each video for semantic adherence and physical commonsense on 1-5, and the joint score counts videos with both at least 4. At v1 the best of seven models, Wan2.1-14B, reaches 32.6 percent overall and 21.9 on the hard split; momentum and mass conservation are the most violated laws (about 40 percent). Its 7B auto-rater improves on Gemini-2.0-Flash-Exp but physical-commonsense correlation is still only 0.37.
• articleHow does PhyGround score physics per law, how reliable is its human study, and how well does its open PhyJudge-9B judge track humans?
  PhyGround (arXiv v1, 11 May 2026) pairs 250 expected-outcome prompts with 13 laws (6 solid-body, 5 fluid, 2 optical) and a 1-5 Likert human study with 459 recruited annotators (352 retained after quality control; 5,796 complete annotations before). Eight models were scored: Wan2.2-27B-A14B and Veo-3.1 tie at 3.28/5 Overall and no model reaches 3.3; solid-body contact laws (momentum, collision) are hardest, and Veo-3.1 leads fluid and optical. The authors release PhyJudge-9B (a LoRA adapter on Qwen3.5-9B); its aggregate relative bias is 3.3 percent versus 16.6 for Gemini-3.1-Pro, measured only on held-out Veo-3.1 videos.
• articleWhat are the code, weights and dataset terms of the main 4D reconstruction and video-to-4D releases, and where do the texts conflict?
  Per-component record at pinned commits. Code: 4DGaussians (Apache text), Real-time 4DGS and DreamGaussian4D (MIT text), Consistent4D (Apache), Dynamic 3D Gaussians (MIT except Inria-licensed files) and SV4D inference code (MIT). Four repositories carry a conflict between the root licence and Inria non-commercial source headers or files (4DGaussians, Real-time 4DGS, DreamGaussian4D) or a stated split (Dynamic 3D Gaussians). Weights: SV4D 2.0 under the Stability AI Community License with a card that contradicts itself at exactly US $1,000,000 revenue; SV4D v1 card unreadable without login; Consistent4D and Dynamic 3D Gaussians advertise third-party or archive checkpoints without terms. Datasets: Neu3D is CC BY-NC 4.0; D-NeRF/HyperNeRF/Nerfies dataset terms are not stated in the READMEs read; DyCheck's iPhone dataset is Apache 2.0 per its README. No legal clearance is concluded.
• articleFor each system that turns structural-control renders into photoreal video, what is actually released (code, weights, licence) and what is closed or unavailable?
  Cosmos-Transfer1 and 2.5: code under Apache-2.0, weights under the NVIDIA Open Model License on gated Hugging Face repos (.pt checkpoints). VACE: Apache-2.0 code and ungated Wan2.1-VACE weights (safetensors) whose repo ships an Apache-2.0 licence text; the LTX variant carries a different licence tag. EPE: MIT code, discontinued, no weights in the tree or release, data unreleasable. RealMaster: paper only, no official code found. Runway Aleph and Luma Modify Video: closed metered services. Licence prose for Cosmos weights is not re-read here (C3-B019-W6 covers it).
• articleWhat do the HunyuanDiT, FlashVDM and bria-rmbg-1.4 agreements require of someone who reuses or redistributes TripoSG code or output: territory, user caps, output-use limits, attribution?
  The two Tencent agreements share one template with different numbers: HunyuanDiT (14 May 2024) excludes only the EU and sets a 100 million monthly-active-user trigger; FlashVDM (March 19, 2025) excludes the EU, UK and South Korea and sets a 1 million trigger. Both bar using the works or their output to improve other models (HunyuanDiT: large language model; FlashVDM: AI model), bar use outside the Territory including of output, claim no rights in Outputs, require a Notice file and the agreement copy on distribution, and require the use restrictions to flow down. The RMBG-1.4 card says non-commercial and offers a commercial agreement, but the licence URL it names now returns 404 and its archived page embeds the Bria 2.0 agreement, which does not name RMBG.
• articleWhy do latent action models learn noise, camera motion or scene cues instead of actions, and which remedies do the papers report?
  Four papers analyse why latents go wrong. Zhang et al. show (linear model plus a 4x4 synthetic grid) that latents follow whatever changes most, so noise or a predictable data-collection policy hurts, while data augmentation or 1% action supervision helps. Lee et al. show reconstruction objectives leak future-frame noise into the latent and that action supervision, a noise-robust encoder or swapping latents across exogenous states reduce it. Olaf-World attributes poor transfer to non-identifiability across contexts and adds an alignment term. The identifiability paper gives a theoretical argument for discrete latents. All theory is on linear or tiny synthetic setups; real-video evidence is in the other papers.
• articleHow do the primary papers evaluate latent actions, and which of those checks could an agent re-use on its own model?
  Every paper uses a different test, so numbers cannot be compared across papers. Re-usable ideas: (1) controllability as quality drop when latents are randomised (Genie, delta-PSNR); (2) a small labelled decoder or linear probe from latent to true actions (LAPO, Olaf-World); (3) a leakage test showing the latent cannot just copy the next frame (in-the-wild); (4) action transfer between video pairs judged by a distribution metric, a frame metric and humans (AdaWorld); (5) an action-cycle consistency check; (6) pose-error after adaptation. Most real-video transfer judging is qualitative or small human panels.
• articleWhat do the learned latent actions actually correspond to, and how can I find out what each one does?
  What a latent corresponds to is an empirical question per dataset. LAPO found, on synthetic Procgen data with known actions, clusters that match the true actions in many games and fragments in games with partial observability. Genie authors report consistent latent meanings (for example down, up, left on robot video) by inspection of generated frames. On diverse in-the-wild video, a later study reports actions that are camera-relative and localised in space, not tied to one body. None of these gives a naming for latent codes in advance; you find out by trying each code.
• articleI need a tiny 3D file to test a glTF loader: which Khronos sample model should I start with and where do I read its licence?
  At glTF-Sample-Assets 15cd02d3, licences are stated per model directory: Box ('One mesh and one material. Start with this.') is CC-BY-4.0 for model files, Triangle is CC0-1.0 for model files with CC-BY-4.0 metadocumentation, and both exclude logos and trademarks; the repository README defers to each model's README for detailed licence information.
• articleWhat exactly does Blender's USD exporter write into UsdSkel joints, bindTransforms, restTransforms and blend shapes?
  At Blender v5.2.2 (d13f752e) the USD exporter writes Skeleton.joints in pre-order bone-tree order but SkelAnimation.joints in pose-channel order, writes bindTransforms from bone arm_mat and restTransforms from the parent-relative pose matrix when a pose channel exists (so the rest attribute reflects the current pose, not necessarily the rest pose), stores bone lengths as a Skeleton primvar, and authors blend shapes without inbetweens in the files read (static reading).
• articleAfter automatic rigging or auto weights, how do I check and fix the skin weights in Blender?
  Problem. Auto weights often deform badly. Conditions. Armature-deformed mesh in Blender. Method. (1) Parenting with Automatic Weights uses a distance-based 'bone heat' calculation; the manual warns it can produce overlaps, with symptoms visible when posing (parts do not deform as expected), and then weights must be edited by hand. (2) Enter Weight Paint Mode and inspect each bone's vertex group: blue is near 0.0, red near 1.0. (3) Weights usually have to be normalized for deformation; use Normalize All and Auto Normalize while painting. (4) Pose the rig through extreme poses (elbow, shoulder, hip, fingers) and look for tearing or collapse. Example. Parent mesh with automatic weights, pose the arms overhead, then paint fixes in problem vertex groups. Failure modes. Dense overlapping limbs and clothing; empty vertex groups for non-deform bones. Limitations. Manual text only; no test on a specific mesh was run here. Version/date. Manual source fetched 2026-10-23.
• articleWhat do Tencent's Hunyuan3D-2 licence and README say about territory, user limits, output rights and PBR, and does this block studio use?
  Problem. A model labelled open-source may carry territorial and user-count conditions. Conditions. Using Hunyuan3D-2 family weights or code. Method. (1) Read the Community License: it is limited to a Territory that excludes the EU, UK and South Korea; use of works or Output outside it is unlicensed. (2) If the licensee's products exceed 1 million monthly active users on the release date, a licence must be requested from Tencent. (3) You must not use the works or Output to improve another AI model; Outputs themselves are not Model Derivatives and Tencent claims no rights in Outputs, but responsibility for Outputs sits with the user. (4) Pipeline: README describes a two-stage design (bare mesh, then texture synthesis) and VRAM of about 6 GB for shape and 16 GB for shape plus texture; the 2.1 release is announced as adding a PBR model, under its own repo and licence (not read here). Failure modes. Using outputs for a European release; training a downstream model on outputs. Limitations...
• articleWhat does TRELLIS output for an image-to-3D asset, and what do its README and licence imply for game or film use?
  Problem. Image-to-3D tools produce meshes of unclear production readiness. Conditions. Using Microsoft TRELLIS to make assets from an image. Method. (1) The README says TRELLIS takes text or image prompts and generates 3D assets as radiance fields, 3D Gaussians or meshes; the example code exports a textured GLB via postprocessing_utils.to_glb with a simplify ratio (0.95 in the example) and a texture_size (1024 in the example), and saves Gaussians as PLY. (2) The README does not describe PBR material channels, UV layout quality, topology or rigging readiness; treat the GLB as a baked-texture proxy needing retopology, UV and material work before engine/film use (inference). (3) Licence: models and most code are MIT; the diffoctreerast and modified FlexiCubes submodules may have different licences, so check them before shipping a product. Failure modes. Assuming MIT covers the whole tree. Limitations. No visual inspection of outputs was performed. Training-data licence questions are co...
• articleCan I use the Inria reference 3D Gaussian Splatting code in a commercial or client project, and which popular alternatives carry different licences?
  Problem. Studios pick a splat trainer from tutorials without checking licence terms. Conditions. Any paid, client or product use of splat training or rendering code. Method. (1) Read LICENSE.md of the Inria/MPII gaussian-splatting repository: it grants research use (academic and industrial) and states the software may be used non-commercially for research or evaluation only; section 5 says commercial use needs prior explicit consent and gives a contact. (2) Treat derivative works as bound by the same use limitation (section 4.2). (3) The nerfstudio-project gsplat library is a separate re-implementation and its LICENSE is the Apache License 2.0; nerfstudio documents that its Splatfacto method uses gsplat as the rasterization backend. (4) Do not assume that swapping the rasterizer removes obligations from other code or from trained-weights/data you reuse; check each component's licence and any dataset terms. Example. A pipeline that shells out to the Inria train.py inherits the non-co...
• articleWhat exactly do the Mesh Cache modifier's Time and Factor modes compute for PC2 and MDD files, and is the PC2 expression ((time / fps) - start) / sampling consistent with the units the modifier and the exporter use?
  Code reading at main commit f0e5cc32 (observed 2026-10-22, not executed). In Scene play mode the modifier sets time = ctime/fps for both Time and Factor, then time = Frame Scale x time - Frame Start. The PC2 Time branch then divides by fps again: frame = ((time/fps) - start)/sampling, whereas the MDD Time branch compares time (seconds) directly with the file's stored per-frame times and ignores fps. Blender's own PC2 exporter and manual define start and sampling in frames, so for exporter-written PC2 the Time branch mixes units unless fps = 1. Factor mode maps clamp(time, 0, 1) x frame_tot to a float index, so the last sample is reached at (N-1)/N and held to 1. No bug report or test mentions this (candidate 01). Whether this is intended for PC2 files from other tools is unverified.
• articleHow do Unity Addressables decide which bundles players re-download after a content update, and what role do the content state file and catalog hash play?
  Per Addressables 2.3 docs, a full rebuild with a new catalog forces players to re-download all remote bundles, whereas a content update build driven by the original release's addressables_content_state.bin keeps unchanged bundles under their original names so only changed bundles download; code changes require a new player build because bundles carry no code; and a catalog hash file beside the catalog (same name, different extension) limits catalog downloads to real changes.
• articleIf I author a USD ParticleField3DGaussianSplat and render with hdPrman (RenderMan), does it read the sorting/projection hints or kernel type, and what does it draw?
  Corrected 2026-10-23 - see C3-B642-06; the original statements below are kept for history and are superseded where the correction says so. At OpenUSD v26.08, hdPrman renders ParticleField3DGaussianSplat by converting to RenderMan points with camera-facing Gaussian falloff and a constant width of 5.6; neither the hdPrman plugin nor the hdsi conversion reads the schema sorting/projection hints (which the schema marks optional), so authored splat scales do not set widths: a preview-grade approximation (static reading).
• articleWhich authored Gaussian-splat attributes carry through to a RenderMan render of USD, and which are lost?
  Corrected 2026-10-23 - see C3-B642-06; the original statements below are kept for history and are superseded where the correction says so. In OpenUSD (ee47c679) renderers without native particle-field support get splats converted to points: the generic hdsi conversion sets point width from the mean of the three scales and turns spherical-harmonic radiance into primvars, while the hdPrman plugin replaces width with a constant 5.6; orientation and kernel are not used in either file, so the RenderMan path is a preview of positions and colour, not a faithful Gaussian splat render.
• articleHow does KHR_mesh_quantization interact with skinning, morph targets and animation, and what must an encoder do?
  Corrected 2026-10-23 - see C3-B642-04; the original statements below are kept for history and are superseded where the correction says so. KHR_mesh_quantization lets POSITION, NORMAL, TANGENT and TEXCOORD (and morph deltas) use byte/short types and adds no new dequantisation transform: non-skinned meshes fold it into the node transform (possibly via an extra leaf node so existing animation is undisturbed), skinned meshes into inverseBindMatrices, and texture coordinates into KHR_texture_transform; signed-normalized decoding differs on OpenGL ES 2.0/WebGL1.
• articleWhich glTF material extensions survive conversion to UsdPreviewSurface and MaterialX gltf_pbr, and what do the converters say they do?
  Corrected 2026-10-23 - see C3-B642-08; the original statements below are kept for history and are superseded where the correction says so. UsdPreviewSurface deliberately omits subsurface and anisotropic specular, so glTF extensions such as KHR_materials_anisotropy have no UsdPreviewSurface input; MaterialX gltf_pbr (nodedef 2.0.1) does carry transmission, specular, IOR, iridescence, sheen, clearcoat, emissive strength, volume attenuation, anisotropy and dispersion. Adobe's USD glTF plugin imports base PBR to UsdPreviewSurface (specular workflow off) through an intermediate ASM model and lists many KHR_materials extensions as supported, without saying UsdPreviewSurface itself carries them.
• articleIs the RenderMan 26 Points primitive documented for splat-sized particles? Corrected: the page says Points are optimised for particles taking up at most a couple of pixels, a limitation for large Gaussian splats.
  Problem. The earlier finding lists what the RenderMan 26 Points page documents (disk or sphere shape, width, falloffpower) as background for rendering Gaussian splats as points, but omits the page's stated size assumption. Conditions. RenderMan 26 public documentation. Method. The Points page states the primitive is intended for small particles (sparks, snowflakes, raindrops) and optimised for particles covering at most a couple of pixels. Failure modes. Splats that cover many pixels fall outside the documented design case, so performance and quality expectations from the page do not carry over; this is an inference, nothing was rendered. Limitations. One page; other RenderMan documents were not searched. Version and date. RenderMan 26 pages fetched 2026-10-23.
• articleDoes the Unity Smart Merge manual say anything about .meta files and which YAML files UnityYAMLMerge will not merge? Yes: it merges .unity, .prefab and .meta files and, by default, does not merge .asset, .mat or .controller files, handing them to a fallback tool.
  Problem. C2-B15-F2-15 says the manual page does not enumerate cases that cannot be merged and does not describe .meta handling. Conditions. Unity 6.6 (6000.6) Smart Merge page. Method. The page states UnityYAMLMerge merges scene, prefab and .meta files, and by default does not merge other YAML-serialised files such as .asset, .mat or .controller, which it passes to a fallback tool from mergespecfile.txt. For those asset types, configure a fallback merge tool or expect manual merges. Failure modes. Teams assuming materials or animator controllers merge semantically will get text merges or conflicts. Limitations. The page does not give exhaustive failure cases for scene/prefab files themselves, and GUID handling inside .meta is not explained. Version/date. Unity 6.6, fetched 2026-10-23.
• articleDoes the Adobe USD glTF plugin still treat UsdPreviewSurface as the main output? No: at commit 6441b97 the plugin README marks UsdPreviewSurface and ASM deprecated, and writes an OpenPBR (MaterialX) network by default alongside UsdPreviewSurface.
  Problem. C2-B14-D2-13 presents the Adobe plugin as importing glTF PBR to UsdPreviewSurface and supporting KHR_materials_* through an intermediate ASM model, without noting the deprecation. Conditions. Adobe USD-Fileformat-plugins main head 6441b97 (2026-07-17), gltf/README.md. Method. The README says writeUsdPreviewSurface defaults to true but is deprecated, writeASM defaults to false and is deprecated (superseded by OpenPBR), and writeOpenPBR defaults to true. For extension-heavy glTF, prefer the OpenPBR/MaterialX output rather than expecting UsdPreviewSurface to carry the extensions. Failure modes. Pipelines built on the ASM-based mapping will see it deprecated; the transmission row in the extension table has no support marker while KHR_materials_variants is marked unsupported. Limitations. Whether OpenPBR carries each extension is not established here. Version/date. Commit 6441b97, still main head on 2026-10-23.
• articleDoes hdPrman always convert USD Gaussian-splat ParticleField prims to RenderMan points? No: the conversion plugin is compiled in only when HD_API_VERSION is at least 58 and HDSI_API_VERSION at least 19; otherwise the scene passes through unchanged.
  Problem. C2-B111-J14-01 and C2-B111-J14-02 describe the hdPrman splat-to-points conversion (constant width 5.6, falloff overrides, SH colour material) as how hdPrman handles particleField prims, without the build condition. Conditions. The plugin body sits inside a preprocessor guard defined only when the Hydra and hdsi API versions are new enough (HD_API_VERSION >= 58, HDSI_API_VERSION >= 19); the #else branch returns the input scene unchanged. Method. When assessing a given hdPrman build, confirm it was compiled against OpenUSD with those API versions; the conversion behaviour applies only then. Failure modes. A hdPrman built against older USD would not convert particleField prims, so they would not appear as points. Limitations. The file is identical in v26.08, v26.11.0, v26.11.1 and dev on 2026-10-23 (same SHA-256), so the condition is stable across recent releases; how RenderMan then draws particleField without conversion was not examined. The overrides values (ri:falloffshape ...
• articleDoes KHR_mesh_quantization allow unnormalised byte/short NORMAL and TANGENT, and is the extension optional? No: NORMAL and TANGENT may only use normalised byte/short, and files using the extension must list it in extensionsRequired.
  Problem. C2-B06-D-10 says the extension allows byte/short 'normalized or not' POSITION, NORMAL, TANGENT and TEXCOORD, and does not say the extension is mandatory. Conditions. glTF assets using KHR_mesh_quantization. Method. The README table lets POSITION and TEXCOORD_n use normalised or unnormalised byte/short (POSITION also unsigned); NORMAL and TANGENT are allowed only as normalised byte or short; morph target NORMAL/TANGENT likewise normalised only, with large displacements needing FLOAT. Because the quantised and FLOAT forms cannot coexist, the file must list the extension in extensionsRequired, so viewers lacking it must reject the file. Failure modes. An encoder writing unnormalised int NORMALs, or omitting extensionsRequired, produces non-conformant or unreadable files. Limitations. Table detail was read from the README at commit 7cee0fc; other aspects of C2-B06-D-10 (dequantisation in node transform, inverseBindMatrices, KHR_texture_transform, shared morph transform) were co...
• articleWhat do Houdini, Maya and Epic documents say about carrying hair/groom curves and widths through Alembic and USD?
  Partially verified: the curator re-read S18, S19, S22; S20, S21 were not re-read by the curator and rest on the producer's pinned reading. Houdini 22.0 imports Alembic curve sets as open polygons (no separate groom type in that mapping); its SOP Import LOP writes cubic-only BasisCurves for Hydra or NurbsCurves for round-tripping, and offers Set Missing Widths because USD's fallback width is 1.0. Unreal's Alembic groom importer reads ICurves, converts DCC width to groom_width without overwriting an authored groom_width, and falls back to 1 cm when no width exists, so missing widths behave differently in USD (1.0 scene unit) and Unreal grooms (1 cm).
• articleHow does Apple say 180, 360 and wide-FOV (fisheye) video is signalled in QuickTime/MP4 for APMP, converted from Google Spherical Video tags, streamed with HLS, and played back in visionOS?
  Partially verified: the curator re-read S3; S4 was not re-read by the curator and rests on the producer's pinned reading. In Apple's Projected Media Profile the projection kind is the discriminator (rectilinear for 2D/3D/spatial, half-equirectangular for 180, equirectangular for 360, parametric for wide-FOV); HLS publishing declares stereo and projection in REQ-VIDEO-LAYOUT and the init map segment must repeat the format information.
• articleHow does the ambientCG API identify assets, signal releases and file changes, express dimensions and state licence, and how stable is it?
  Partially verified: the curator re-read S19, S20, S27; S21, S22, S23, S24, S25, S26 were not re-read by the curator and rest on the producer's pinned reading. ambientCG states all assets are CC0 1.0 with no credit needed, but its API records carry no licence field (licence comes from the policy page), sampled dimension fields are 0 with no unit defined, and the operator warns the API may not be stable enough for enterprise use; v3 (from 2026-02-28) and repackaging updates change file names.
• articleWhich VBench-2.0 dimensions can I run on my own clips (for example previs shots), what do they need, and which only work on the standard prompt suite?
  At the pinned repo commit, only five of the 18 dimensions accept custom videos: Human Anatomy, Human Identity, Human Clothes, Diversity (needs 20 clips per prompt) and Multi-View Consistency. The other 13 (including Camera Motion, Instance Preservation, Motion Rationality and all Controllability and Physics state-change dimensions) assert-fail on custom input and need the standard prompt suite, which means sampling the model under test on fixed prompts (3,860 videos for a full run by my count). Requirements are heavy (CUDA 11.8, flash-attention build, several third-party weights, some from Google Drive/Dropbox).
• articleWhat does VBench-2.0 actually measure (five categories, 18 dimensions), how is each dimension scored, and how is a single Total Score built?
  VBench-2.0 (arXiv v2) scores generated video on 18 dimensions grouped into Human Fidelity (3), Creativity (2), Controllability (7), Physics (4) and Commonsense (2). Scoring mixes generalist judges (LLaVA-Video-7B captions or yes/no questions, Qwen2.5-7B-Instruct as text judge) with specialist detectors (ViT human-anomaly detectors, ArcFace, CoTracker heuristics, a Qwen2.5-VL-3B LoRA anomaly detector). The paper reports per-dimension scores for four systems and no aggregate; the equal-weight Total Score is defined only in the repository README and used by the leaderboard code.
• articleWhich results from planning-with-a-learned-model papers could transfer to media or agent workflows, and which are specific to robotics or games?
  In the inspected papers the demonstrated uses are games, simulated control and robot manipulation; none reports a media-production or LLM-agent workflow. What may transfer is method-level (plan by sampling sequences through a learned predictor, check model exploitation by comparing virtual versus real score, use a latent rather than pixel predictor for speed) not the results. Specifically, V-JEPA 2-AC needed a goal image and a reward-free robot dataset; Dreamer and TD-MPC2 need reward signals; the exploit-the-dream failure from 2018 is the most portable lesson. This is a synthesis; no experiment was run.
• articleWhat terms and scope do the public TD-MPC2 code, checkpoints and datasets state, and what is left open?
  TD-MPC2 has the clearest statements of the three planning systems: the repository README says the project is MIT-licensed (LICENSE: MIT, Copyright Nicklas Hansen 2023) with third-party code under its own licences; the Hugging Face model card says in prose '**License:** MIT' for the 324 checkpoints (12 multi-task from 1M to 317M, 312 single-task at 5M), and the dataset card says MIT for the 30-task and 80-task replay-buffer datasets (545M and 345M transitions). The card's out-of-scope note says checkpoints are not expected to generalise to unseen tasks. Not stated: redistribution of simulator-derived data, model_training on outputs, and the terms of the underlying simulators (DMControl, Meta-World, ManiSkill2, MyoSuite) and of MuJoCo assets.
• articleWhat control signal does GameNGen condition on, and what resolution, context and speed does the paper state?
  GameNGen's ICLR version (arXiv v2) says the action set is 'key presses' of DOOM and that each action is embedded into a single token; the paper text never lists the individual actions (no hit for mouse, keyboard or discrete in the whole PDF text), so the exact action set is not stated. Stated numbers: 320x240 frames padded to 320x256, context of 64 frames plus 64 actions (a little over 3 seconds), 20 fps with four DDIM steps on a TPU-v5 (50 fps with distillation), trained on RL-agent play where each agent action is held for 4 frames. Table 2 shows PSNR rising from 20.94 at one history frame to 22.36 at 64.
• articleBeginner: what minimum evidence should an agent require before telling a production that a world model 'worked' on a job, and which found cases meet it?
  PROPOSAL (synthesis, not a tested standard): require eight items before saying a world model worked on a job: system and version, production date, task, inputs, outputs used in the final cut, a comparison baseline that was actually run, who measured and who wrote the claim, and what was not measured. Mapping the cases found: none of the generative cases meets items 1, 5 or 6; the only case with a run alternative (DOSENMUTTER) is not a world-model case. Until a case meets the checklist, an agent should say 'vendor reports' and not 'it worked'.
• articleWhat acceptance check could an agent run before relying on a streaming video or world model for a long take (proposal, not a tested protocol)?
  No source gives a ready acceptance test. Combining measures the papers use gives a proposal: generate at the longest length you need, score per-segment brightness/imaging quality and a motion measure, compute first-vs-last delta, run a copy-first-clip control, and separately test whether the model returns to an earlier place (memory), which the papers say these methods do not provide. This is synthesis, untested.
• articleHow do COLMAP, Meshroom, Metashape and RealityScan establish real-world scale and units for a photogrammetry result?
  Partially verified: the curator re-read S15, S19; S24, S25, S26 were not re-read by the curator and rest on the producer's pinned reading. Reconstructions have no inherent real-world scale: COLMAP geo-registers by fitting a similarity transform to known camera-centre positions, and Meshroom's SfMTransform can scale to real-world size (its auto_from_landmarks mode only fits landmarks into a [-1,1] box, i.e. normalises rather than measures). RealityScan's one-distance rule and Metashape's markers/scale bars are producer-read only.
• articleWhat capture rules do COLMAP, Meshroom and Agisoft Metashape state for photos used to build a 3D asset, and where do their numbers differ?
  Partially verified: the curator re-read S14, S18; S15, S22, S24 were not re-read by the curator and rest on the producer's pinned reading. COLMAP asks for high visual overlap with each object seen in at least 3 images; Meshroom asks for 60% side and 80% frontal overlap, no flash, fixed focal length and no shadows, reflections or transparent objects. Metashape's overlap figures (60/80% aerial; 60-70% close range) and its no-crop/rotate rule are producer-read only.
• articleWhat files and 3D-asset outputs (point cloud, mesh, UVs, textures, formats) do COLMAP, Meshroom and Metashape state they produce?
  Partially verified: the curator re-read S14, S15, S20, S21; S16, S24 were not re-read by the curator and rest on the producer's pinned reading. COLMAP's automatic reconstruction writes a fused.ply dense cloud and Poisson and Delaunay PLY meshes, with texturing as a separate mesh_texturer step producing an atlas with per-face UVs; Meshroom's Texturing writes texturedMesh.obj/.mtl and texture_*.png, prefers Alembic for point-cloud interchange and has FBX support paused. Metashape's export list is producer-read only.
• articleFrom USD render settings to EXR frames, what do UsdRender and OpenEXR actually define for AOV names, layers, parts and ID mattes, and which choices are left to the renderer?
  Partially verified: the curator re-read none of the cited sources; S21, S22, S24, S25 rest on the producer's pinned reading. UsdRender lets a RenderProduct list RenderVars by name and source but USD does not standardise RenderVar names, and OpenEXR channel and layer names carry no stacking or compositing meaning (multipart names must be unique; colorInteropID now carries colour space); AOV names and recombination therefore depend on a renderer-specific naming table passed to compositing.
• articleWhat do the UsdSkel schema docs require for joint order, bind/rest transforms and inbetweens, and how do Maya-USD and Houdini KineFX treat them?
  The OpenUSD v26.08 UsdSkel overview defines bindTransforms as world-space and restTransforms as local-space (both in joints order; rest optional only with a complete non-sparse animation), the schema supports inbetween blend shapes, and maya-usd (cfdac964) writes them with weight-derived names, unit-scaled offsets and normal offsets.
• articleHow machine-readable is per-asset licence and credit metadata in the Poly Haven API, the Sketchfab Data API, OpenGameArt pages and Kenney pages, and what must an importer parse out of free text?
  Poly Haven's API gives per-asset authors and descriptions but no licence field (CC0 is stated site-wide; only 3 of 521 model descriptions mention it), so an agent must take the licence from site terms. Sketchfab exposes a structured licence object on model detail, but unfiltered search results for non-downloadable models carry a null licence, so licence presence depends on downloadability and endpoint. Author credit is a structured field only on these two services among those surveyed.
• articleHow does Godot decide that an imported asset in .godot/imported is stale and must be reimported?
  Godot (code at 6790667) uses a cheap mtime test (source and .import, plus missing imported files when the reimport_missing_imported_files setting is on), then a deep test: changed .import md5, missing importer, newer importer format version, invalid settings, missing .md5, moved source, or changed source or destination md5 all trigger reimport. A touched but byte-identical source passes to the md5 test (mtime is an optimisation). The docs describe only the source-md5 trigger and erased .godot/imported files, and recommend not committing .godot/.
• articleWhat do Houdini Karma and Blender Cycles documentation say about naming, compressing and handing AOVs (including Cryptomatte) to compositing?
  Partially verified: the curator re-read S31, S33; S32, S34, S35 were not re-read by the curator and rest on the producer's pinned reading. Karma lets multipart EXR compression be set per AOV via a Render Var LOP and its Cryptomatte vars take an overlap limit (6 recommended) and an external manifest (embedded as metadata if unset); Blender shader AOV names are free text used in both the AOV Output and Render Layers nodes, provided they do not clash with enabled passes, and Cycles light groups give per-light Combined passes.
• articleWhich AOV names and settings does Hydra itself standardise for render delegates, and how are render settings changes tracked?
  Partially verified: the curator re-read S25, S26; S24, S27 were not re-read by the curator and rest on the producer's pinned reading. Hydra names standard AOVs by tokens (e.g. color with pre-multiplied alpha, depth, cameraDepth, primId) plus primvars:/lpe:/shader: prefixes (shader format marked TBD), and describes each AOV by format, multiSampled, clearValue and a settings map. These tokens are a separate naming layer from USD RenderVar names and EXR channel names.
• articleWhich glTF extensions does Godot 4.7.2-stable register, what happens to unsupported required extensions, and what do docs say about blend shapes and animation import?
  Partially verified: the curator re-read GL1, GL2, GL6; GL3, GL4, GL5 were not re-read by the curator and rest on the producer's pinned reading. Godot 4.7.2 imports a glTF only if every extensionsRequired entry is in its supported set: eight extensions built into GLTFDocument plus those declared by registered document extensions (physics, KTX for KHR_texture_basisu, WebP); otherwise it errors suggesting a missing GLTFDocumentExtension plugin. gltf_document.cpp contains no Draco, meshopt, quantization or advanced PBR material extension names (curator string search of that file only); plugins can extend the set.
• articleIf I author projectionModeHint or sortingModeHint on a USD gaussian splat, will it change how the Pixar-shipped renderers draw it?
  Partially verified: the curator re-read S1, S5; S6, S7, S8 were not re-read by the curator and rest on the producer's pinned reading. At v26.11.1 nothing changes from v26.11.0 in the files read: hints are passed to Hydra but the sample splat renderer still sorts by view-space depth. Treat the hints as metadata for other renderers, not as a control of look or correctness in Pixar's in-repo code.
• articleFor Alembic exports headed to Unreal, how do velocities, changing topology, sub-frames and face sets get written and consumed?
  Partially verified: the curator re-read none of the cited sources; S11, S24, S25, S26, S5 rest on the producer's pinned reading. For Alembic caches into Unreal: if topology changes, export per-vertex velocities and use them as motion vectors (otherwise let Unreal compute motion vectors), and export face sets or no materials are created; Blender writes velocities only from a point attribute named 'velocity' and writes face sets when the option is on; Maya's open-source AbcExport offers -writeFaceSets but the file read contains no velocity export.

More