3D and worlds

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

• articleIn an OpenEXR handed to compositing, how are layers, parts and channel names defined, and what does Blender's writer do when it writes single-part layers versus multipart?
  OpenEXR names layered channels L.C (nestable) without any stacking or compositing meaning; multipart parts need unique names (dots allowed) and may differ in compression and data window and mix deep and flat. Blender v5.2.2 writes either single-part channels layerpass.view.C or, in multipart mode, one part named layerpass-view with channels layerpass.C.
• articleHow does UsdRender describe render settings, products and AOVs (RenderVars), and what does it leave undefined (naming, file layout)?
  UsdRender (v26.08) describes outputs as RenderSettings -> RenderProduct (raster or deepRaster; productName passed as authored to the driver) -> RenderVar (prim name is the variable name; sourceType raw/primvar/lpe/intrinsic, with intrinsic unimplemented and LPE extensions non-portable). USD sets no universal RenderVar names, and per-product camera/resolution overrides may force separate renders.
• articleHow do Alembic point caches (OPoints/IPoints) differ from PC2/MDD in identity, topology change and time sampling, and what does a mismatch look like?
  Alembic point samples carry positions with optional uint64 ids, velocities and widths; the API comments describe ids on the first sample and re-stated ids/topology to mark a topology change, and time is per-property TimeSampling (Uniform, Cyclic or Acyclic). The writer enforces positions on sample 0 but substitutes an empty ids array, so readers cannot assume ids exist. PC2/MDD carry neither ids nor topology.
• articleWhat exact binary layouts and checks does Blender's Mesh Cache modifier apply to PC2 and MDD files (header, vertex count, frame count, endianness)?
  Blender's PC2 reader expects a 32-byte native-endian header ('POINTCACHE2' + file_version, verts_tot, start, sampling, frame_tot) and MDD a big-endian pair (frame_tot, verts_tot) followed by per-frame float times; both refuse a vertex-count mismatch or frame_tot <= 0, and neither checks vertex order, which the manual says must stay consistent.
• articleDo Godot USD add-ons exist that can write USD for offline finishing, and what do their READMEs state about coverage and builds?
  Of two third-party Godot USD GDExtensions read, only godot-usd documents writing USD (scene-graph authoring save with warnings for unsupported edits; prebuilt binaries for macOS arm64 only); godot-usd-importer documents import only. Neither documents rendered-frame output.
• articleWhich Unity Recorder 5.1.7 image-sequence settings matter for handing real-time renders to offline finishing (format, bit depth, colour space, alpha)?
  For scene-linear handoff, Unity Recorder 5.1.7 documents a narrow combination (HDRP, non-Game-View source, EXR, tonemapping and dithering off) for 'Linear, sRGB (unclamped)'; EXR is written at 16 bits; alpha is unavailable from Game View or non-alpha formats.
• articleWhat does the Unity USD package document for exporting scenes or animation to USD, and what are its stated limits?
  The Unity USD package (3.0.0-exp.5, experimental) exports selected GameObjects and transform overrides to USDC/USDA/USDZ, recommends the Recorder route when runtime compatibility is not needed, and is pinned to USD v20.08 with no Apple Silicon native editor; it is a scene-graph export, not rendered frames.
• articleAt Godot 4.7.2-stable, how do resource UIDs (.uid sidecars, uid in .tscn/.tres headers and in .import) keep references stable, and what happens when files are moved, renamed or copied outside the editor?
  At Godot 4.7.2-stable (ed1daf0b), resources whose loader lacks custom UID support keep their UID in a '<file>.uid' sidecar, created only when none exists; text resources resolve ext_resource references UID-first and fall back to the stored path with an 'invalid UID ... using text path instead' warning; the UID registry is cached in uid_cache.bin under the project data path; the class docs state the sidecar rule.
• articleHow do the MPEG glTF extensions (MPEG_media, MPEG_buffer_circular, MPEG_accessor_timed, MPEG_texture_video, MPEG_scene_dynamic, MPEG_viewport_recommended) wire timed video, V3C and dynamic scene data into a glTF scene?
  The Khronos-hosted MPEG glTF extension READMEs (7cee0fcd) are based on ISO/IEC FDIS 23090-14 and define timed media through MPEG_media (alternatives with required mimeType, listed value video/mp4), a circular buffer (count default 2, minimum 2), a timed accessor that scenes with timed data shall use (suggested update rate default 25) and a video texture (format default RGB); each README states Copyright ISO/IEC 2022 with use subject to ISO's licence and IPR policy, not the repository licence.
• articleHow does OpenUSD UsdVol represent OpenVDB volumes and their animation over time?
  At OpenUSD v26.08 (ee47c679), a Volume binds fields through 'field'-prefixed relationships, an OpenVDBAsset field is animated by authoring filePath time samples per frame (the overview example maps frames 101-104 to separate .vdb files) because '$F' pattern substitution is not supported, fieldDataType must be authored (a missing value is an error) and fieldClass maps to OpenVDB grid classes; the same schema file defines ParticleField prims.
• articleIs the 1X world-model challenge data and are the GENIE_138M/35M weights gated or behind a terms-acceptance step on Hugging Face, and what licence label does each repository carry?
  At observation (2026-10-22 13:29-13:30Z) the Hugging Face API returned gated:false for all four repositories (tokenised dataset, raw-video dataset, GENIE_138M, GENIE_35M) to unauthenticated GETs; no terms were accepted or needed for the metadata. Licence labels: tokenised data apache-2.0; raw video cc-by-nc-sa-4.0; both GENIE weight repositories apache-2.0; the 1xgpt GitHub repository carries a full Apache-2.0 LICENSE text. Only labels (not licence prose) are present in the Hugging Face repositories, and the dataset ids used in older links now redirect.
• articleIf I run MEt3R, which reconstruction backbone do the paper and the repository say it uses, and which one do I get by default?
  The paper text (arXiv v1 and v2) names DUSt3R as MEt3R's reconstruction model and says MASt3R was not used. The released code at commit ee0e175 defaults to a 'mast3r' backbone (MASt3R weights) and offers 'dust3r' and 'raft' as options, and the README overview still says DUSt3R. So a default run is not the configuration the paper text describes; to match the paper wording pass backbone="dust3r". Whether the paper's tables were produced with DUSt3R weights is not stated by any source read (see candidate 02).
• articleWhat exactly do Wayve's GAIA-2, GAIA-3 and GAIA-4 pages claim about fidelity, controllability and correlation with real driving, and which of those claims come with a value?
  Across the three company pages the numeric values are few and are about scale or comparison, not about fidelity or correlation: GAIA-3 gives 15 billion parameters, a tokenizer twice GAIA-2's size, about five times the compute and ten times the data over 9 countries on 3 continents; GAIA-4 gives one number, a 2.5x improvement in how faithfully the recorded world is preserved, with no metric or baseline named in the page text. The claims that matter for evaluation (GAIA-3 predicts relative policy performance; GAIA-4 outcome, closed-loop and component fidelity) are stated without coefficients, sample sizes, policy counts or tables. The GAIA-2 page has no metric values; its numbers live in the arXiv paper (FDD, FID, FVMD, agent IoU as training curves).
• articleWhich licence or terms text covers each of the five 3D asset sources behind TRELLIS-500K (Objaverse-XL, ABO, 3D-FUTURE, HSSD, Toys4k), and what does each text say about commercial use, redistribution and training?
  The five sources carry five different kinds of terms, and none of the five texts mentions model training. ABO is CC BY 4.0 as a whole. Objaverse-XL is ODC-By 1.0 as a whole with the licence of each object left to the object. 3D-FUTURE (agreement of March 28, 2020) is research-only with bars on commercialisation and on handing the data to third parties. HSSD is CC BY-NC 4.0 (project page) and gated; its paper says the Floorplanner data is made available for academic use. Toys4k has no licence file in the repository: it asks users not to distribute the data and refers to a separate attribution PDF that was not read. This is a reading of the texts, not a legal conclusion.
• articlePoly Haven, ambientCG and Kenney all say CC0: what do their own pages actually say about commercial use, redistribution, modification, attribution and AI training, and which site-level terms sit outside the CC0 asset licence?
  At the access time below, the three libraries each state CC0 for their assets and say commercial use needs no credit. Only Poly Haven addresses AI training in its own words (FAQ: yes). Poly Haven also keeps separate website and API terms that are not the CC0 asset licence, and its FAQ states a "re-license" limit that the CC0 legal code does not contain. ambientCG and Kenney pages read say nothing about AI/ML training; that is a bounded absence, not permission or prohibition.
• articleExactly how does the Matrix-Game GameWorld Score measure keyboard and mouse following, what do Oasis, MineWorld and Matrix-Game 1.0/2.0 score on it, and what does it need to run?
  GameWorld Score (Matrix-Game 1.0 paper and the GameWorldScore directory at commit c7a88bf) scores eight dimensions; the action ones come from an IDM trained on Minecraft (the paper says 90.6% keyboard accuracy, R2 0.97 for mouse) applied to generated clips. The code asks for 76 actions on 32 initial images, i.e. 2432 videos. Paper results (author-run): keyboard/mouse accuracy 0.77/0.56 Oasis, 0.86/0.64 MineWorld, 0.95/0.95 Matrix-Game 1.0. Matrix-Game 2.0 reuses the score on Minecraft (0.91/0.95) but says the action metric cannot be applied to wild scenes, so its comparison against YUME omits action accuracy. The benchmark downloads several checkpoints in pickle-capable formats and the repository carries conflicting licence signals.
• articleWhat does HOOD, an open learned cloth simulator trained without a ground-truth solver, report about quality, speed and limits compared with MeshGraphNets-style and physics-based cloth?
  HOOD (CVPR 2023, arXiv v3) combines graph networks with hierarchical message passing and trains self-supervised with a physics-based loss instead of solver-generated targets. Its human study (30 participants, pairwise realism) favours HOOD over two learned baselines and is about equal to the ArcSim physical simulator on the sequences where ArcSim could run; HOOD runs at about 7-14 fps for a 12K-vertex dress ablation (not a comparison with ArcSim time). Stated limits: no garment-garment interaction (self-penetration), failure when body motion exceeds training velocities, and wrong correspondences under severe body self-intersection. Code is MIT-licensed per the LICENSE text; data and SMPL models are fetched separately.
• articleWhat do V-JEPA 2 and V-JEPA 2-AC checkpoints come with in code, weights and outputs terms, and what does the public text leave open?
  At commit 204698b45b3712590f06245fbfba32d3be539812 the README says 'the majority of V-JEPA 2 is licensed under MIT' with three Apache-2.0 source files and says nothing about checkpoints. Weights are hosted two ways: dl.fbaipublicfiles.com URLs listed in the README (including V-JEPA 2-AC, a 11.76 GB object; HEAD only, not downloaded) and 8 Hugging Face repos whose card front-matter says 'mit' for six and 'apache-2.0' for the two ViT-g fpc64 pretrained cards; the cards contain no licence prose and the HF repos list no LICENSE file. V-JEPA 2-AC is not in the HF collection. The paper (arXiv v1) says it trained V-JEPA 2-AC on under 62 hours of Droid robot video and pretrained on public datasets including YT-Temporal-1B; rights of that data are not addressed in the texts read. Terms for outputs, redistribution and training use of the weights: not stated.
• articleIs WHAM's context 10 seconds or 1 second, and do the two statements count different units?
  They count different things, and the card contradicts itself. The Nature paper's methods say the 1.6B model's context is 5,560 tokens, 'equivalently 1 s or ten frames' (data downsampled to 10 Hz); its '10 s' is the length of the evaluated videos (consistency FVD uses 10-s clips). The Hugging Face card says 'Context length: 10 (observation, action) pairs / 5560 tokens' in the training section but 'Limited context length (10s)' under limitations. Ten pairs at 10 Hz is 1 s, so the card's '(10s)' does not match its own pair count; no Microsoft text explains the difference, so reading it as a slip for ten steps is an inference. Code defaults agree with ten frames: run_dreaming.py sets context_length 10.
• articleWhat exactly does Microsoft's WHAM/Muse release offer for gameplay ideation, and how do the Nature paper and the model card differ?
  WHAM is a single-game model (Bleeding Edge), 300x180 at 10 Hz, with a 1-second context window, built from a user study of game creatives and evaluated on consistency (FVD), diversity (Wasserstein) and persistency (blinded human labels). The released 200M and 1.6B checkpoints are research-licensed (non-commercial, no distribution), not real time, and domain-bound. The model card's 'Limited context length (10s)' conflicts with the paper's 1 s of context; the card's own '10 pairs' agrees with the paper.
• articleDoes the Alembic project publish sample .abc files, and what terms cover them?
  The Alembic repository does not publish a curated sample-scene set. Its 46 .abc files are test inputs: mostly fuzzer and malformed-input regressions and security proof-of-concept files, plus two small renderer test files (prman cube.abc and xforms.abc). The only licence file is a BSD 3-clause text addressed to source and binary forms, which does not mention the .abc data, so reuse terms for those files are unclear.
• articleWhat is the AOUSD USD Core Specification, which document is the published version, and where do I read it (and under what licence)?
  The only published AOUSD Core release in the repository is v1.0.1 (2025-12-12, PDF and Markdown), which AOUSD's announcement calls Core Specification 1.0. It is CC BY-ND 4.0: verbatim copies with attribution may be redistributed, but distributed adaptations may not.
• articleHow do the original 3DGS PLY layout, Khronos KHR_gaussian_splatting and Niantic SPZ differ in coordinate system, colour (spherical harmonics), activation (scale/opacity) and rights, so that conversions between them are correct?
  The reference 3DGS code (54c035f7) writes PLY columns holding pre-activation parameters (log scale, pre-sigmoid opacity), so copying them into KHR_gaussian_splatting (linear scale, 0..1 opacity, degrees 0-3, required colour space) needs exp/sigmoid conversion; SPZ (MIT) stores RUB coordinates, names PLY as typically RDF and GLB as LUF, and supports SH degree 4, which KHR cannot carry without truncation. The three components carry different terms: 3DGS code research/non-commercial, SPZ MIT, KHR text Khronos copyright with a CC-BY-4.0 header.
• articleWhich public OpenUSD test assets exist (usd-wg/assets, Pixar Kitchen Set), and which of them state terms that would let L10 use them?
  The usd-wg/assets repository (3b75c2da) declares licences per asset in each readme (typically CC0 or highly permissive) with the repository licence as fallback, while Pixar's Kitchen Set is offered only under a temporary, revocable licence for personal, non-commercial testing of USD.
• articleWhich public museum open-access programs release images for reuse, and which parts (data, images) are actually under CC0?
  The Met and NGA open-data READMEs state that image files are not part of their CC0 datasets, while Cleveland states images for up to 30,000 works are offered on the same terms as its dataset; all three ask users not to use museum trademarks or imply endorsement. Dataset-level CC0 therefore does not establish image rights for Met or NGA items.
• articleWhich glTF-Sample-Assets models carry non-CC0/CC-BY terms (non-commercial, EULA, testing-only) that L10 must exclude or handle separately?
  At 15cd02d3 the repository's REUSE.toml has 151 annotation blocks: 62 plain CC0-1.0 and 46 plain CC-BY-4.0 (curator re-tally matches the producer); others combine CC licences with trademark references or name non-open terms (SCEA, CRYENGINE agreement, 3DRT testing-only, Poser EULA, Adobe Stock, CC-BY-NC-SA-4.0, CC-BY-4.0 AND CC-BY-NC-4.0). VirtualCity's 3DRT grant is for testing glTF tools only, and DamagedHelmet lists CC-BY-4.0 and CC-BY-NC-4.0 together without file-level scope.
• articleWhat files, metadata and HLS settings does Apple state for authoring and publishing Apple Immersive Video, and where do the stated bitrate tiers disagree?
  Apple describes Apple Immersive Video as captured at 90 fps with more than 50 MP per eye, targeting 60 pixels per degree and up to 230 degrees field of view; WWDC25 session 403 recommends MV-HEVC 4320x4320 per eye at 90 fps in P3-D65-PQ for HLS, with the AIME VenueDescriptor and APAC audio referenced from the playlist. The session's spoken tiers (25-100 Mbps average, 50-150 Mbps peak) conflict with its own summary (25-150 Mbps average); the conflict is preserved.
• articleWhere do licence statements, dependency notices, dataset provenance and weight file formats disagree or leave gaps for these 3D generators?
  Four items need curator attention: (1) TripoSG advertises MIT but its NOTICE lists code derived from Tencent HunyuanDiT (Tencent Hunyuan Community License notice) and the BRIA RMBG-1.4 background-remover whose card says non-commercial; the README downloads RMBG weights. (2) TRELLIS-500K metadata says MIT while the DATASET.md describes assets curated from Objaverse(XL), ABO, 3D-FUTURE, HSSD and Toys4k and states the dataset is CSV metadata; asset-level licences were not read. (3) Hunyuan3D repos ship .ckpt (pickle-capable) next to safetensors; 2.1 ships .ckpt and .bin; TripoSR ships model.ckpt; the licence texts do not address file formats. (4) The Tencent Hunyuan 2.0 HF card links its licence at LICENSE.txt while the HF file list shows LICENSE.
• articleWhich systems claim to steer a running world model or long video with text, how was that steering actually tested, and what did the tests show?
  Four measured protocols for text steering exist in 2025-2026 preprints, plus one system (Zing-0.5) that shows it only qualitatively, and each measures something different. Hunyuan-GameCraft-2 uses a VLM judge on six interaction dimensions over 100 test images against image-to-video baselines. Incantation uses blinded human ratings (three annotators, 400 clips) of whether an entity performed the prompted action. LongLive uses CLIP score per 10 s segment over 160 videos with six prompts each. WorldSonus compares a mid-stream sound prompt switch against a no-switch hold with identical visuals (audio only). Zing-0.5 measures only navigation (WBench) and shows text events as a qualitative session whose lasting consequences the authors say are not established. No source tests a text event in a released interactive model with persistence over later actions.
• articleFor the audio-capable world models and sync-focused sound tools found here, what do the repositories and model cards actually say about licence and weights per component, and where do they disagree?
  HelixWorld (Noiz AI) is the one audio-visual world model with a published checkpoint: code and the HF card say Apache-2.0, but the repository NOTICE says the LTX-2 code and model assets inside it are governed by the LTX-2 Community License (paid licence for entities with at least US$10M annual revenue; fine-tuned weights count as derivatives), the paper says HelixWorld builds on LTX-2.3, and the Gemma text encoder has its own terms. WorldSonus is CC BY-NC 4.0 with .pt weights (pickle-capable). V-AURA code is MIT but its checkpoint is a tar.gz with no stated weight terms. Kling-Foley publishes a paper and a CC-BY-4.0-labelled benchmark dataset but no code or weights were found.
• articleWhat do Fab and Quixel say a buyer of a Fab Standard License or CC-BY asset may do, and what part of the binding text could not be read without a browser?
  The Fab EULA page (Internet Archive snapshot 2026-10-08) shows only a self-described non-binding summary: commercial use and distribution of Projects allowed, stand-alone resale or free redistribution not allowed, no credit to the publisher required. It does not mention AI or machine learning. The binding EULA text is script-rendered and was not captured, so the Standard License's AI-training position is unknown to this batch. Fab documentation separately says the NoAI tag marks assets not to be used for generative-AI data collection and needs the Standard licence.
• articleFor a Sketchfab model, which document governs what I may do (the per-model Creative Commons label, the Standard or Editorial licence, or the Terms of Use), and how do the NoAI and CreatedWithAI labels bear on use in a production or an AI dataset?
  Sketchfab layers three documents: a per-model licence label exposed by the API (seven CC types plus free-st, st, ed), the paid Standard/Editorial licence agreement, and Terms of Use effective 12 August 2026 that bar using NoAI-tagged content in generative-AI datasets and say the licence in force at download governs. Sketchfab's own 2023 blog says the NoAI tag may not bind outside the platform for CC-licensed models. The API objects inspected show no NoAI field, and the licence agreement page text read has no NoAI clause although the blog says Standard-licensed downloads are contractually covered; these gaps are for curation.
• articleWhich of the scene-graph systems have released code, under what stated licence, and what third-party models, keys or data does each need?
  At the observed commits, Hydra (BSD-style LICENSE), ConceptGraphs (MIT) have permissive code licences, while the 3DSceneGraph repository carries a non-commercial research-use software licence and its data sits behind a separate click-through agreement that was not accessed. ConceptGraphs code is MIT but running it needs third-party segmentation and vision-language checkpoints plus an OpenAI API key, whose terms are separate. A permissive code licence does not state anything about weights, data or outputs.
• articleHow does WorldMem store and retrieve earlier frames so a generated Minecraft or RealEstate10K world stays consistent after the camera leaves and returns, what revisit numbers does it report, and what does it cost?
  WorldMem keeps every generated frame as a token-level memory unit with a 5-number pose and a timestamp, picks up to a handful of units per step by field-of-view overlap and time difference (greedy, with a similarity filter), and lets the generating tokens cross-attend to them with pose/time embeddings. Its revisit tests compare regenerated frames against ground-truth frames with PSNR, LPIPS and reconstruction FID: 100 frames generated after a 600-frame memory bank on a custom Minecraft set, and five start=end trajectories over 100 RealEstate10K scenes. These are paper_result numbers on simulator poses; the paper's own tables disagree with each other and with the repository's expected-results block on some cells.
• articleWhat does WorldMemBench's fixed-reference multi-revisit protocol measure, what did it find about decay across returns, axis switches and off-screen events, and how reliable are its checks?
  WorldMemBench (260 cases: 200 static with human, object and environment targets, 60 dynamic with off-screen events) compares every prescribed return with the same first-visit reference over translation loops and yaw/pitch cycles, giving a score per revisit and counting a missing target as zero. On nine open video world models, the static leader (HY-WorldPlay) initiates only 31.7% of prescribed events, scores decay from the first to the third return on all eight metrics, and switching from a yaw cycle to a pitch cycle costs more. The paper adds pose-qualified subsets to separate memory from bad camera execution, and a context-policy study on LingBot World v2. Human agreement is 78.5-92.5% for static metrics but model-level correlation varies from 0.301 to 0.988.
• articleWhich of the explicit-memory world-model papers and revisit benchmarks can I actually download or run, and what do their licence texts, metadata and gates say per component?
  Observed on 2026-10-22: WorldMem has code (non-commercial licence text) and public ungated checkpoints with no licence field; SPMem has Apache-2.0 code and Apache-labelled checkpoints; VMem has MIT-labelled code that vendors a CC BY-NC-SA component and weights behind a Hugging Face access gate; Context-as-Memory has a released dataset but no code I could find; WorldTrace has no code I could find. The two newest benchmarks (R2M-Bench, WorldMemBench) release MIT code and data, but WorldMemBench's dataset card says its third-party images and the generated videos keep their original rights, and its examples set is labelled CC BY-NC by metadata. Nothing here was run, so 'released' means files are listed, not that they work.
• articleHow does VMem index past views with surfels, what is its cycle-trajectory revisit test, and which parts of its release are open, gated or conflicting?
  VMem keeps the past views as images but indexes them through a coarse surfel model: each surfel stores which views saw it, new target cameras render the surfels to vote for the top-K most relevant past views, and an image-set generator (Stable Virtual Camera, SEVA) generates the next views from those. Its main contribution to evaluation is a cycle-trajectory protocol (go out, return along the same path, score every 10 frames of the return) with LPIPS/PSNR/SSIM plus rotation and translation distance of camera poses re-estimated from the generated frames. The tables contain rows that are identical across methods, which should be checked before reuse. Code is MIT-labelled but vendors a CC BY-NC-SA component, and the weights sit behind a Hugging Face access gate.
• articleIf I run the open Oasis 500M code, what resolution, frame rate, context length and action vector does it actually use, and which of those does the project text state?
  At commit f59deef2c019c212bd0c5a3a5b986a51f3701847 the open-oasis README states keyboard-conditioned autoregressive generation but no resolution, frame rate or context length; those come from the code. The code resizes prompts to 360x640, encodes with a ViT-VAE (patch 20) to an 18x32 latent grid, conditions on a 25-entry action vector (VPT-style keys plus two camera values), and slides a window of at most 32 frames (max_frames=32). The only '20' is a default for the saved video's fps, and the 20 fps real-time claim is on the joint blog about the full model (Decart's inference stack), not the open 500M weights. The weights repo is gated and was not opened.
• articleDo the Cosmos repositories, model cards and docs restate the NVIDIA Open Model License accurately, and do the dated versions that appear in model-card front matter differ from the current text?
  Three dated agreement texts are readable without accepting anything: the current page (October 24, 2025), the text embedded in the Cosmos-Predict2.5-14B card (September 23, 2025, an abridged rewording), and the text embedded in the Cosmos-Tokenize1 card (January 6, 2025). The January text uses 'Model Derivatives', has no Special-Purpose Model clause, no 'Guardrail' definition and no Separate Components section, and its guardrail sentence has no 'substantially similar Guardrail' exception. GitHub READMEs and the docs licence page restate the licence only by name and link. One Cosmos 3 model repository ships a LICENSE file that matches the October 2025 text. Model-card 'Important Note' paraphrases of the guardrail clause differ between cards.
• articleBefore I use, host, fine-tune or redistribute a Cosmos model that is under the NVIDIA Open Model License, what does each clause of the current agreement text say, and which of my questions does it not answer?
  At the dated text 'Last Modified / Version Release Date: October 24, 2025' the agreement has 11 numbered sections. It grants a perpetual, worldwide, no-charge, revocable licence to use, modify, make derivatives of, sell and distribute the Model; allows redistribution with a copy of the agreement and a Notice line; requires 'Built on NVIDIA Cosmos' for Cosmos Models in several uses (including training another AI model with them or their outputs); states NVIDIA claims no ownership of outputs; and ends your rights automatically on litigation or on bypassing a Guardrail without a substantially similar one. It has no clause on hosting as a service by that name, no express patent licence (the word 'patent' appears only in the litigation-termination sentence) and no operative commercial-use clause beyond the preamble statement. This is a reading of text, not legal advice.
• articleThe agreement ends my rights if I bypass a Guardrail without a substantially similar one. What do the agreement text, the Cosmos docs and the Cosmos code at a pin say about whether the guardrails can be turned off?
  Three source kinds disagree on tone, not necessarily on fact: the agreement (October 2025) ends rights if you bypass a Guardrail without a substantially similar one; the Cosmos docs state the Cosmos-1.0-Guardrail model 'cannot be disabled' in the diffusion and autoregressive pipelines; and cosmos-predict1 at one commit defines a --disable_guardrail command-line argument and forwards it to the pipeline. The docs sentence and the code are about overlapping but not proven-identical pipelines; the code path was not traced or run. 'Substantially similar Guardrail appropriate for your use case' is not defined in the text.
• articleMy stereo or immersive video file is going to Apple Vision Pro: which Apple movie profile (3D, Spatial Video, APMP for delivery, APMP for production) applies, and what does each allow?
  Apple's Movie Profiles for Spatial and Immersive Media (v0.9 Beta, 9 June 2025) distinguishes 3D/Spatial Video (rectilinear MV-HEVC) from APMP delivery and production; APMP for production permits other codecs and frame-packed video while delivery is MV-HEVC, with vexu boxes signalling projection, packing and disparity adjustment.
• articleI have splat values from a typical training export (log scales, logit opacity): which values must I convert before putting them in a glTF KHR_gaussian_splatting file or a USD ParticleField3DGaussianSplat prim?
  Both glTF KHR_gaussian_splatting and OpenUSD ParticleField3DGaussianSplat store linear scale and linear 0..1 opacity; the OpenUSD example converter applies exp() to PLY log scales and a logistic sigmoid to PLY logit opacity and treats rot_0 as the quaternion real part; glTF marks out-of-range opacity invalid and defines diffuse colour from SH0 with a +0.5 bias at render time.
• articleWhich third-party parts does the 1X tokenised-data repository actually contain, as opposed to merely reference, and what terms statement exists for each component?
  By its own listing the repository contains one third-party-credited checkpoint (magvit2.ckpt, 1.15 GB), the 1X decoder script, and the token and state files. The Cosmos Tokenizer is referenced, not bundled: the listing has no .jit or Cosmos files, and the 1X decoder script tells the user to download the NVIDIA repository and weights first and loads decoder.jit from a local pretrained_ckpts path. Per component: tokens, script and ckpt sit under one repository-level Apache-2.0 line; Open-MAGVIT2's own files are Apache-labelled but the bundled file matches none of them; the Cosmos weights are under the NVIDIA Open Model License (card-linked June 2024 text), a separate text. No two non-empty texts disagree about the same component, so no conflict is filed; the holds are listed.
• articleWorldMark reports a 'Response Latency' for ten released interactive world models: what does it measure, is it compute or wall-clock latency, what are its values, and how do they relate to throughput figures?
  WorldMark v2 (Alaya Lab and others; not the model authors) scores 'Response Latency' as how soon the generated video's optical-flow speed reaches half its normal sustained speed after a command switch, normalised to 0-100 (higher is better). It is a content-response metric, not compute time: the paper reports no generation time or FPS, only that scoring one model on one split takes about one hour on two H200 GPUs excluding video generation. First-Person Real translational/rotational Response Latency: Matrix-Game 2.0 98.23/86.47, LingBot-World 86.56/99.20, HY-World 1.5 85.44/85.13, HY-GameCraft 1.0 90.97/51.99, Yume 1.5 40.66/77.11. Yume 1.5 has the lowest translational Response Latency here yet the highest Hz in WorldRoamBench, so the two measure different things. The v1 HTML has six models and no latency metric.
• articleWhat non-author evidence exists on the time from starting an interactive world model to the first usable frame, which parts of that cold-start time are separable (loading, compile or kernel warm-up, offload transfers), and which were fixed upstream?
  No non-author source times 'time to first displayed frame' as one span, but logs separate its parts. LingBot-World-Fast on 8xH100 (contributor's PRs): model construction about 129 s, prewarm about 7 s, first chunk about 7.2 s versus about 1.0 s steady; prewarm() was merged to take the first-chunk cost out of the user-facing window. HY-WorldPlay Wan-5B on 4xA100 (user): first four chunks 89-368 s, then about 3.4 s; a maintainer attributes it to torch.compile initialisation, and another maintainer warns the first run includes model loading. Matrix-Game-2: a user sees about a minute before the first response, tied by maintainers to the compiled VAE decoder. Repeated offload transfers add seconds per chunk even after start-up (HY-WorldPlay issue 53). Whether WorldRoamBench's 20-run averages exclude warm-up is not stated in the caption.
• articleWhich released interactive world models have a speed or latency figure measured by someone other than their authors, where is it, and which systems have none?
  Non-author timings exist for several systems but live in only four kinds of place: a benchmark's efficiency table (WorldRoamBench v5 times 10 open models on one H20), an efficiency paper's baseline row (QuantWM v3 times LingBot-World-v2 and HY-World 1.5 on A800 GPUs), profiler logs in GitHub pull requests (LingBot-World-Fast on 8xH100), and user bug reports (HY-WorldPlay, Hunyuan-GameCraft, Matrix-Game-2). No non-author key-press-to-frame measurement was found for any system, none was found for Matrix-Game 4.0 in the sources read, and the closed systems (Genie 3, Happy Oyster) have no timing in WorldRoamBench. WorldMark's 'latency' is an in-video response metric, not wall-clock time.
• articleWhere have learned models been documented for line work in a drawn-look pipeline (production vs research vs open tools), and what are their inputs, limits and terms?
  Documented production use: Sony Pictures Imageworks trained a scikit-learn GradientBoostingRegressor inside Houdini to predict how artist-drawn ink lines should adjust with camera angle and expression, ~600 frames per character first generation, with artists nudging the prediction (account by Grochola via fxguide, 2019). This model moves hand-drawn lines on 3D animation; it does not extract lines from images. Research/open line extractors differ: Informative Drawings (arXiv 2203.12691v3) maps photographs to line drawings, with a stated dependence on a pretrained depth network; ControlNet 1.1 lineart models use extracted or hand-drawn lines to steer image generation. None of the two research/open sources documents use on 3D character renders.
• articleWhich controls and evaluations from driving and robot world models could carry over to media previs or virtual camera planning, and what has not been shown?
  SYNTHESIS, not a measured result. Transferable pieces: (1) camera-pose conditioning from a single image with Plucker rays (Cosmos; joystick-style paths shown qualitatively, scored by pose re-estimation on static scenes); (2) multi-view synchronised re-rendering along arbitrary camera trajectories (Predict2.5 robot multiview; cross-view Sampson error 19.73 px vs 26.61 px baseline); (3) re-driving a logged scene while holding everything else fixed ("world-on-rails", GAIA-3/4 company pages) and embodiment (rig) transfer; (4) edit-by-partial-noising for weather/time-of-day and box-guided agent inpainting (GAIA-2, qualitative); (5) language -> trajectories (DriveDreamer-2). Transferable evaluations: re-estimate the camera path from the output and compare with the commanded path (pose success, rotation/translation error, Vista IDM), Sampson cross-view error with a real-footage floor. Not shown by any source read: film-style subjects (actors, close-ups), cuts, lens language, or non-static ...
• articleHow do robot world-model sources define action conditioning and what do they measure (video quality, value prediction, policy ranking), and which numbers support the claim that a world model can rank robot policies?
  Robot world models condition on low-level action vectors: a 7-D gripper delta on Bridge (Cosmos), whole-body command sequences (1XWM). Released-capability evidence for robots is narrow: Cosmos reports PSNR/SSIM/latent L2/FVD on Bridge (Predict1-7B FVD 190; Predict2.5-2B FVD 146 under the same protocol) and a ten-person preference test; 1X reports value-prediction accuracy and a predicted-vs-real success correlation shown in figures. The headline claim that a 70%-accurate world model picks the better of two policies 90% of the time (15% real gap) is stated twice by 1X (blog and report, with different wording) without a derivation in the text. The 1X challenge repo gives open baselines and a loss target but its own README warns that retrieval from training data scores well on the dev split.

More