3D and worlds

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

• articleWhich paid provider models can run inside ComfyUI as built-in "Partner" (API) nodes, and how does that change workflow design?
  Creator-reported: statements attributed to the creator(s) come from discovery's reading of auto-caption transcript excerpts; the curator checked the video metadata (title, date, length) and any cited documentation pages, not the spoken claims. Nothing was tested; promotional or sponsored framing applies where stated. The ComfyUI docs index (captured 2026-10-23) lists 298 built-in Partner node pages: Video 126, Image 101, 3D 45, Audio 16, Text 10. By name prefix the largest groups are Tripo (24), Recraft (20), Bria (19), Kling (19), ByteDance (18, i.e. Seedance/Seedream), Luma (15), Vidu (15), Flux/BFL (12), Wan API (12, incl. Wan 3 reference/image-to-video), Gemini (10), MiniMax (9), PixVerse (9), Meshy (8), ElevenLabs (8), Rodin (7), HeyGen (7), Grok (7), OpenAI (7), Runway (7), Tencent (6), Ideogram (6), plus Magnific, Fish, Topaz, Krea, Reve, LTX, Sync, Beeble, Veo, Claude and Higgsfield entries. Implication: a single ComfyUI graph can mix local open-weight models with hosted clo...
• articleHow do you "AI-render" a 3D layout locally (3D passes -> video model), and how did the best local approach change from Feb to Aug 2026?
  Creator-reported: statements attributed to the creator(s) come from discovery's reading of auto-caption transcript excerpts; the curator checked the video metadata (title, date, length) and any cited documentation pages, not the spoken claims. Nothing was tested; promotional or sponsored framing applies where stated. Three Mickmumpitz videos trace one pipeline across 2026. Feb (V1): Inner-Reflections' VACE + SkyReels V3 R2V merge (Apache-2.0 card) reads control passes (VACE) and reference images (SkyReels); export an outline pass (Blender Workbench flat black + white outlines, or Freestyle as a render pass) and a normalized/inverted Z depth pass with a contrast curve; outlines keep composition but let the model invent detail, depth follows geometry more tightly, blend both if needed; keep control-net strength low for start-frame image generation (Z-Image Turbo + control nets); free workflow ~120 frames at 24 fps; paid version chains 81-frame chunks with 11 overlapping start frames f...
• articleHow does 4D Anyone create a 4D Gaussian-splat capture of a person from one phone video, and what are its limits in a ComfyUI port?
  Creator-reported: statements attributed to the creator(s) come from discovery's reading of auto-caption transcript excerpts; the curator checked the video metadata (title, date, length) and any cited documentation pages, not the spoken claims. Nothing was tested; promotional or sponsored framing applies where stated. Mickmumpitz (2026-09-11) ports the 4D Anyone paper (ant-research, Apache-2.0 repo) to ComfyUI: from one (ideally tripod) vertical phone video, a 3D skeleton locks pose per frame and a fine-tuned video model generates up to ~48 consistent synthetic views (it keeps already-generated angles in context and reshuffles which views generate together: far-apart first for shape, neighbours later for smooth transitions), then gsplat trains per-frame Gaussian splats. He swapped the original skeleton code (not licensed for commercial production) for Meta SAM 3D Body (SAM License) and bundled a gsplat setup script. Settings: camera preset (high-quality vs quick 16-view), Turbo on (l...
• articleHow can you generate an explorable 3D Gaussian-splat world from one image or prompt locally, and which world-generation approaches failed on consumer hardware (per Mickmumpitz, Aug 2026)?
  Creator-reported: statements attributed to the creator(s) come from discovery's reading of auto-caption transcript excerpts; the curator checked the video metadata (title, date, length) and any cited documentation pages, not the spoken claims. Nothing was tested; promotional or sponsored framing applies where stated. Mickmumpitz (2026-08-25) reports weeks of R&D: Apple SHARP (single-image to splat in one pass) stitched across 360-degree slices produced seams and fell apart off-axis, and its licence did not suit commercial users; Insta360 UniSarp (360 to 3D) degrades quickly when moving; HunyuanWorld 2.0 needs two large models resident and targets 4 GPUs; NVIDIA Lyra ran but needs 91 GB of checkpoints, Linux-only, 6-minute first start; World Labs Marble is fast and crisp but restricts movement area and lacks working reflections. Chosen base: Skywork Matrix-3D (MIT): 360 panorama -> depth (MoGe) -> rough mesh -> render camera paths -> Wan 2.1 video model + a lightweight LoRA in-paints...
• articleWhich metrics, baselines and benchmark protocols do the 4D generation papers use, and which reported numbers can be compared across papers without being misled?
  Partially verified: the curator re-read none of the cited sources; S1, S3, S8, S13, S11, S5, S6, S15, S16 rest on the producer's pinned reading. The papers use different metric families and protocols: CLIP R-Precision plus human votes (MAV3D), CLIP Score plus pairwise user study (4D-fy), LPIPS/CLIP/FVD on a seven-object synthetic benchmark (Consistent4D, DreamGaussian4D, L4GM, SV4D), CLIP-F/CLIP-O and image metrics on a 20-case held-out set (Diffusion4D), and view/frame FVD variants plus FV4D (SV4D family). Cross-paper tables carry copied or inconsistent rows: the DreamGaussian4D row in L4GM and SV4D tables (0.16/0.87) equals the Consistent4D row and differs from DreamGaussian4D own Table 3 (0.12/0.92); Consistent4D FVD is 1133.44 in two papers and 1133.93 in a third; L4GM calls the benchmark eight animations while Consistent4D and its uid file list seven. User-study percentages point in opposite directions in MAV3D and 4D-fy. The one independent-protocol re-run found (SV4D 2.0 on i...
• articleCan I get an animated file I can open in Blender, Unreal or Unity (mesh sequence, PLY sequence, glTF, Alembic) out of these text/image/video-to-4D systems, and what did the authors actually release?
  Partially verified: the curator re-read none of the cited sources; S1, S3, S2, S12, S11, S6, S7, S5, S17, S18, S9, S4, S14 rest on the producer's pinned reading. Among the eight systems, only DreamGaussian4D documents a command-level path to a DCC tool: its README exports a per-frame OBJ mesh sequence and loads it in Blender with the Stop-motion-OBJ add-on. MAV3D describes a mesh-sequence conversion in the paper but has no available code. L4GM ships weights and code whose inference script writes only mp4 renders; a PLY writer exists in its repository but is not called, so a per-frame PLY sequence is an untested route. SV4D / SV4D 2.0 releases sample view-by-frame image grids, and no fitting or export code is visible by path name. No source read mentions Alembic, glTF, FBX or USD output.
• articleWhat do text-to-4D, image-to-4D and video-to-4D generation mean in practice, and for MAV3D, 4D-fy, DreamGaussian4D, Consistent4D, SV4D/SV4D 2.0, L4GM and Diffusion4D, what goes in, what representation comes out, and is it something I can run?
  Partially verified: the curator re-read none of the cited sources; S1, S3, S11, S13, S15, S5, S8, S4, S6, S12, S17, S9 rest on the producer's pinned reading. "4D" here means a 3D object or scene that changes over time and can be rendered from new viewpoints; every system read produces a generated asset, not a captured one. Three routes appear in the primary sources: (a) per-asset optimisation against diffusion priors, by score distillation (MAV3D, 4D-fy, Consistent4D) or, for DreamGaussian4D, by fitting Gaussian deformation to a generated driving video with image-diffusion guidance, (b) a diffusion model first generates a grid of views by frames, then a 4D representation is fitted to it (SV4D, SV4D 2.0, Diffusion4D), and (c) a feed-forward network that reconstructs per-frame 3D Gaussians from a video (L4GM). Output representations are dynamic NeRFs or Gaussians; only the DreamGaussian4D abstract mentions animated meshes (per-frame, candidate 02) and no abstract describes a rigged as...
• articleIs there a dated, machine-readable source for the Physics-IQ Verified leaderboard that gives i2v and v2v scores per model?
  Partially verified: the curator re-read none of the cited sources; S3, S2, S4, S6, S7, S10, S11 rest on the producer's pinned reading. No JSON or CSV results file exists in the benchmark repository at the pinned commit and the leaderboard site exposes no JSON endpoint. Two public sources can be parsed without a browser: the README markdown table at a pinned commit (53 Verified rows, dated, with PR or report links), and typed records embedded in the site JavaScript chunk (52 records with sub-scores, run values, sampling and dateAdded). They agree on 50 of 53 rows within 0.06 points, differ on date-added for most rows and on three Physical Registry and Magi-1 rows, so a curator should cite the README at a commit and treat the site chunk as a convenience copy.
• articleDoes arXiv 2610.06637 show that its state-tracking model actually uses the action sequence and resists a false fact in its state, and how strong is that evidence?
  Two controlled tests in S2, both only on the synthetic CEO-Bench simulator. (1) A fabricated -$5,000/week cash shock injected into the encoded state costs latent-recursive 11.1 SE-score points on average (8-14 at every horizon), latent-SR after distillation 5.4 and the full latent-SR 1.7; an irrelevant latency distractor costs about zero. (2) In a three-branch counterfactual set (60 companies, separate simulator run, horizons 1/3/7/10, statistics at h>=5) only the full latent-SR has a significantly positive adjusted margin (+0.024, sign and permutation p < 0.001); raw-history models are significantly negative on the sign test and every earlier latent-SR stage is null. Neither test exists for Jericho or ScienceWorld, so claims about fiction state tracking from these tests are extrapolations.
• articleIf I want to turn a generated Gaussian-splat world into a playable prototype, what steps did VIVERSE and the Gaussian Mansion developer use and what do their pages say about the result?
  Two World Labs case pages describe game prototypes built on Marble splats. VIVERSE used a five-step flow (generate or select a splat in Marble, import into PlayCanvas, add colliders/triggers/lighting, optimise for browser and mobile, publish). Of three demos only Clockwork Conspiracy was newly generated, from a quoted text prompt. Gaussian Mansion (Unreal Engine 5 rail shooter) used Marble PLY exports stitched in Marble Studio and imported with a 3D Gaussians plugin, and names transparency, material and artifact problems. The only quantity is one developer's "under an hour" quote; neither page reports counts of levels, iterations or ratings.
• articleUnity Manual (R1): IK works backwards from a goal so the end effector reaches it, useful "if you want a character to grab an object or stand on an uneven" surface; API: SetIKPositionWeight, SetIKRotationWeight, SetIKPosition, SetIKRotation, SetLookAtPosition, bodyPosition, bodyRotation; setup:...
  Unity Manual (R1): IK works backwards from a goal so the end effector reaches it, useful "if you want a character to grab an object or stand on an uneven" surface; API: SetIKPositionWeight, SetIKRotationWeight, SetIKPosition, SetIKRotation, SetLookAtPosition, bodyPosition, bodyRotation; setup: valid humanoid Avatar, Animator Controller, enable IK Pass on the layer (which sends OnAnimatorIK), then a script sets goals and weights each frame. R2: animation states with humanoid clips expose "Foot IK" ("Enable to respect Foot IK for this state"). Compared with Unreal's documented plant/unplant thresholds and pelvis limits (C3-B567-01), Unity's built-in layer provides the IK solve and weights but leaves contact detection (raycasts), lock/release hysteresis and pelvis lowering to user code or add-on packages (inference from the API surface; Animation Rigging package not read).
• articleAnswers "how do procedural systems keep feet in contact and the body supported?" from engine documentation. Plant (contact) settings (R1): speed_threshold - "Bone is considered planted below this speed" (speed from an FK speed curve or graph root motion); unalignment_speed_threshold - speed at...
  Answers "how do procedural systems keep feet in contact and the body supported?" from engine documentation. Plant (contact) settings (R1): speed_threshold - "Bone is considered planted below this speed" (speed from an FK speed curve or graph root motion); unalignment_speed_threshold - speed at which the foot becomes fully unplanted, the range between the two approximating the foot's roll phase (the docs themselves note this "feels innaccurate most of the time"); distance_to_ground - within this distance of the plant plane the bone is planted and fully aligned; unplant_radius and unplant_angle - linear and angular deviation that release the lock; replant_radius_ratio/replant_angle_ratio - below these fractions of the unplant limits the foot replants (hysteresis); min/max_extension_ratio bound how far the leg chain may extend; separating_distance keeps the feet apart; ankle_twist_reduction limits slope alignment twist; adjust_heel_before_planting; reconstruct_world_plant_from_velocity...
• articleAt Godot 4.7.2-stable, what sequence of checks over .import files, .godot/imported cache files and recorded hashes decides whether an asset is reimported, and which edits trigger a reimport?
  At Godot 4.7.2-stable (tag ref -> ed1daf0b) the reimport decision is layered: modification-time comparison during scan only queues a test; the test checks the .import md5, then import metadata, then source/dest md5 from a separate .md5 file; the cache base name is keyed on the md5 of the res:// path; docs (godot-docs 9adca4c1) say to commit .import files and that erasing .godot/imported triggers reimport (static reading).
• articleWhat pricing tiers and rate limits does Odyssey publish for its world-model API, and where does the public record stop?
  No price list, per-hour price or named paid tier was found on the public documentation, launch posts or legal page. What is public: a "Free tier" example (5 concurrent streams), plan-based limits shown only in a login-gated dashboard, HTTP 429 for concurrency and simulation-queue limits, and contract terms (fees set in an Order, month-to-month, renewal increase capped at 10% a year, rate limits "as defined by Company from time to time"). The only dollar figure for Odyssey-3 is a vendor GPU-hour assumption for a benchmark cost axis, which is not a price.
• articleWhich failure modes of text world models did arXiv 2604.13824 actually measure (over-optimistic simulation, time-outs, room-connectivity errors), and how large were they?
  S1 measures one failure mode quantitatively: the "too easy simulator" in which a world model lets a weak agent succeed that fails in the real environment (TextWorld, Qwen3-0.6B: 85/200 false positives with the supervised model, 19/200 after BehR; Qwen3-1.7B: 69 and 34). It also reports, descriptively, that 82-97% of WebShop world-model episodes hit the 50-step cap because the model does not reproduce purchase completion, and one worked TextWorld trajectory in which broken room connectivity traps the agent. It reports no per-category error rates for inventory, location or object state.
• articleWhat exactly did arXiv 2604.13824 (v2) test and measure: environments, models, horizon, metrics, headline numbers, limits, and what is released?
  arXiv 2604.13824 (Huang et al.; v1 submitted 2026-04-15 12:56:45 UTC, v2 2026-09-02 08:58:38 UTC, cs.LG, 20 pages) fine-tunes Word2World text world models (Qwen2.5-7B and Llama-3.1-8B bases) with GRPO using BehR, then tests them on WebShop and TextWorld with 200 held-out tasks per domain and four evaluation agents (Qwen3-8B, Qwen3-32B, GPT-4o, GPT-5). Trajectories average about ten steps (cap 50), so this is behaviour at short horizons, not long-horizon modelling. Headline: BehR beats its supervised starting point on CR_pw in 13 of 16 cells with 3 ties; the largest cell is WebShop Qwen3-8B 0.345 -> 0.483 (10/29 -> 14/29 tasks); single-step exact match is preserved or improved in three of four settings (lower for WebShop on the Llama base); WebShop lookahead success rises from 14.5% (ReAct) to 24.0% (W2W-WM) and 25.5% (BehR-WM) for Qwen3-8B. Each configuration was trained once and no seed variance is reported.
• articleWhat is the difference between "state consistency" and "behaviour consistency" for a text world model, and which of the two papers (arXiv 2604.13824, 2610.06637) defines which notion?
  Both IDs resolve to the described papers. arXiv 2604.13824v2 (Huang et al.) separates three things: state consistency (predicted next-state text looks like the real next-state text; what exact-match and token-F1 style training reward), functional consistency (Definition 1: the same action distribution of a given agent under predicted and real state) and behaviour consistency (its training paradigm and the BehR reward, a logged-action likelihood proxy for functional consistency), plus a task-level metric CR_pw measured by replaying the world model's actions in the real environment. arXiv 2610.06637v1 (Wang et al.) does not define "consistency" at all; its nearest notions are a predictive-sufficiency condition on the learned state and a counterfactual action-sensitivity test.
• articleIs there a peer-reviewed study of AI-generated 3D environments for previs with filmmaking professionals, and what does its abstract say was assessed?
  Leininger, Weber and Rothe (IMX 2025, pp. 60-78) report a pilot study with 15 filmmaking professionals comparing three AI-generated 3D environment types (Depth Mesh, 360 degree Panoramic Meshes, Gaussian Splatting) in an Unreal Engine 5 prototype called EnVisualAIzer. The abstract says splats suited high-fidelity previsualisation and panoramic meshes suited rapid ideation, and names limited interactivity and customisation. The abstract states no time saving. Only the abstract (via an index record) was read; the full text could not be read within the rules.
• articleWhich dated, attributable shoots or prototypes used an AI-generated 3D world or scene, and for each one what number does the source actually state and who stated it?
  Seven jobs were read first-hand (pages fetched as raw bytes). Six state a quantity, but every quantity is a company or named-person statement with no described measurement method: over ten environments in one afternoon (Moonshine/VIVE), hundreds of worlds over a few weeks and weeks-to-days (Pixomondo/World Labs launch video), months to about an hour (Veles/Volinga), 15 set-ups over five days (Quite Brilliant, image-based backgrounds, not 3D worlds), ten environments (Promise, Final Twist; no time figure), under an hour for a prototype (one VIVERSE developer). Gaussian Mansion states no number. No source gives a baseline that was actually run, a version, or a quality rating.
• articleWhich Odyssey statements about Odyssey-2 Pro, Max and Odyssey-3 are documented capabilities, which are vendor benchmark claims with stated conditions, and which are marketing wording?
  A scoped ledger. Documented capability (SDK docs): text/image input, 150 s streams, 429 limits. Benchmark claims with conditions: Odyssey-3 Pro 66.1 on Physics-IQ Verified is the best-of-8 entry (the prompt-enhanced entry is 63.4); Max's "highest physics score" uses a vendor-chosen comparator set and partly the vendor's own measurements, and the table caption claims superiority over bidirectional video models that the next sentence says were excluded. Marketing or roadmap: "can run for minutes", robot and driving demonstrations (data sizes but no success rates in the text read). The blog's "quality settings" and "target simulation length" for Simulations are not fields in the documented options object.
• articleWhat resolution, frame rate and latency does Odyssey state for Odyssey-2 Pro and Odyssey-3, and which figures are documented API behaviour versus marketing or benchmark settings?
  For Odyssey-2 Pro the docs give 1280x704 / 704x1280 (with "may vary by model") and a 50 ms frame interval; the launch blog says 720P at 22 FPS. These do not state the same quantities (704 vs 720 lines; 50 ms = 20 per second vs 22 FPS) and no measurement protocol is given. For Odyssey-3 the only resolutions found are benchmark settings (832x480 and 1280x720); no frame rate or latency appears on either Odyssey-3 launch page (searched in extracted text).
• articleHow long can an Odyssey interactive session run, and what concurrency, idle, queue and quota limits does the documentation state?
  The docs state per-stream 150 s and per-connection 60 min defaults, a 15-minute idle timeout, plan-dependent concurrent streams (Free example: 5), a 10-job simulation queue returning 429 when full, and non-refreshable 10-minute browser tokens. A restart after the 150 s lease costs a documented 1-2 s visual gap unless two clients roll over. Marketing wording ("can run for minutes", "120+ seconds of generation") is looser than these figures. All limits are for Odyssey-2 Pro; no limit for Odyssey-3 or Max was found.
• articleWhat input types does the documented Odyssey API accept (text, image, actions) and what does it return?
  In the JavaScript SDK pages read, input is a text prompt plus an optional start image (25 MB cap, resized to 1280x704 or 704x1280) and further text "interact" prompts; scripted simulations add timestamps. No keyboard, gamepad or camera-pose action type appears in those pages. Outputs are a WebRTC video stream, optional broadcast to spectators, and per-stream recordings (MP4, JSONL event log, JPEG thumbnail, preview MP4). A default prompt-expansion step is documented.
• articleHow does a developer actually get access to Odyssey Systems models (Odyssey-2 Pro, Odyssey-2 Max, Odyssey-3), and which of them does the self-serve developer API document?
  At the captures read (2026-10-22), Odyssey documents a self-serve developer API (JavaScript and Python SDKs, API key from a developer portal) only for Odyssey-2 Pro. Odyssey-2 Max is stated to be in private beta for selected partners, and Odyssey-3 is a research preview in a signed-in app plus a contact route. The "Get API Access" link on the Odyssey-3 launch page leads to the same developer portal, whose documentation index names Odyssey-2 Pro only, so no documented API for Odyssey-3 was found.
• articleThe 1X tokenised-data card says 'License: Apache-2.0' and links ./LICENSE; how do I check whether that file exists in the repository listing, and what can I say if it does not?
  At dataset revision 42e3e12f (observed 2026-10-22) the card carries an Apache-2.0 label in its front matter and a closing line linking ./LICENSE. The saved API listing (1331 entries; reread without fetching) has no path matching licence, notice or copying in any case, and the six root files are .gitattributes, README.md, cosmos_video_decoder.py, magvit2.ckpt, unpack_data_test.py and unpack_data_train_val.py. So the linked file has no matching entry in that listing, which agrees with the curator's report. What remains is a label without a licence text in the listing: a rights hold, not a conflict. The check is bounded to an API file listing, not a directory-tree call.
• articleWhat baseline (BF16) generation times does the QuantWM paper report for LingBot-World-v2 and HY-World 1.5, how were they measured, and can they be compared with the authors' real-time claims or with WorldRoamBench?
  QuantWM v3 (Harbin Institute of Technology Shenzhen / NUS; not the model authors) reproduces released models with their official repositories on NVIDIA A800 80GB GPUs and reports BF16 end-to-end generation latency of 193.56 s for LingBot-World-v2 and 112.17 s for HY-World 1.5 (Table 7), with the main evaluation at 480p and 720p, 93 frames. That is 0.48 and 0.829 frames/s (own arithmetic), far below real time, but the Table 7 text does not give GPU count, resolution, checkpoint (distilled few-step or not) or step count, so it cannot be set against the authors' real-time statements. At WorldRoamBench's single-H20 Hz the same 93 frames would take about 52.0 s (LingBot-World) and 71.5 s (HY-World 1.5); the ratios 3.73 and 1.57 are unreconciled.
• articleWhat do users report when they time HY-WorldPlay, Hunyuan-GameCraft and Matrix-Game-2 themselves (including the offloading, compile and consumer-GPU cases), and how do those logs compare with the maintainers' answers?
  User logs show large, flag-dependent gaps. HY-WorldPlay AR-distilled (480p, 125 frames, 4 GPUs): 3-4 minutes reported on 4xH800 and 4xH20 versus about 35 s measured by a maintainer on 4xH20; a user says turning offloading off gives about 35 s (about 3.57 frames/s), and removing repeated transformer CPU-GPU moves cut chunk 1 from 33.59 s to 5.89 s in another report (maintainers merged changes). The Wan2.2-5B variant on 4xA100 shows 90-368 s for the first four chunks and about 3.4 s per chunk afterwards (about 4.71 frames/s if a chunk is 16 frames); a maintainer blames torch.compile start-up. Maintainers say the code is not real time on an RTX 4090. Hunyuan-GameCraft users report about 12 minutes for one action on a 4090 (low-VRAM script) and 40 minutes for under 60 s of video on 8xH20; a maintainer says 8xH20 takes 2-3 times longer than their 8xH800 setting. Matrix-Game-2 users report about a minute of stall before the first response; no user FPS figures were found.
• articleHow do 3D productions get the limited-animation (stepped, "on twos") feel, and what rationale and settings do primary sources document?
  The Xrd developer account (2015) is the primary rationale found: they tried full animation early, it did not read as 2D, so they removed key interpolation (every frame is a key), used about 500 bones per character, avoided simulation, and deformed the mesh on every key to add imperfection. Blender documents a Stepped Interpolation F-curve modifier (step size, offset, start/end) that holds values every N frames. No inspected source documents "on twos" frame numbers for a film; secondary reporting about Spider-Verse exists but no primary source was read (see negative results).
• articleHow do anime-look 3D productions edit vertex normals to get clean face shading, which methods exist, and what do they break?
  The Xrd developers state they edited normals on every major feature and hand-crafted face normals, and did not use normal maps. A Japanese ASW lecture deck (published 2025-07-30; glossed here) names three control methods: place vertices so automatic normals come out right, transfer normals from another mesh, and set normals by hand (reserved for faces; whole-body is called impractical), and lists limits: normals rotate with bones, no subdivision surfaces, extra care with hull outlines. Blender documents Normal Edit (with Max Angle to avoid shading artefacts, described as useful for toon-like shading) and Data Transfer (custom normals).
• articleWhat goes wrong with inverted-hull outlines in production (width vs distance and FOV, hard edges, custom normals, floating lines, fabric) and what fixes did a studio document?
  The ASW English deck (published 2025-11-28; 79 slides) lists concrete failure modes and fixes for the hull: scale width with camera distance and FOV (tan(FOV/2) factor, FOV from the projection matrix), control width with a vertex-colour channel (default 0.5), push the back-face mesh away from the camera to keep only the outer rim, use a second normal set for the hull when hard edges or customised normals break it (stored in the tangent channel at import), and set the width to 0 in valleys to close gaps that make lines float. Unity Toon Shader exposes parallel controls (width map, Z offset, distance fade, baked normal map for outlines). The deck says it avoids code and is engine-independent; the second-normal-in-tangent approach needs an engine import change.
• articleWhich outline method should I start with (inverted hull, post-process edge detection, Freestyle, Line Art), and what does each trade away?
  A practitioner comparison from Arc System Works (speaker Motomura, deck published 2025-11-28) sets post-process outlines (constant line width, lines at mesh intersections, little per-part control) against the inverted hull (per-area control, no intersection lines, thickness upkeep). Xrd used the hull for per-vertex control and in-DCC preview. Unity Toon Shader implements two hull-style modes (extrude along normals, or scale). Blender offers Freestyle (render-time) and Line Art (stroke-based). Epic's post-process documentation shows a CustomDepth/stencil silhouette; it is a general doc with no toon guidance. These are statements by tool makers and one studio; no source compares the methods on equal footing.
• articleIn Blender, what are the line options for a hand-drawn look (Freestyle vs Line Art with Grease Pencil), what does each produce, and what limits does the manual state?
  At Blender 4.5 LTS (manual path observed 2026-10-22), Freestyle is a render-time line engine that outputs lines as a render pass to be composited with Alpha Over, with documented limits (memory, no face-intersection edges, no Z depth). The Line Art modifier generates Grease Pencil strokes from a camera view, can be baked over a frame range, and has option-level fixes (overscan, chaining) for edge continuity. Shader To RGB (EEVEE only) plus Color Ramp is the manual's toon-shading hint. The manual pages read never mention an inverted-hull outline.
• articleHow does toon/cel shading actually work (a stepped ramp on lighting), and what are the first controls to set in Blender, Unity or a custom shader?
  Three primary sources describe the same core: lighting is reduced to a hard step. Motomura (Arc System Works, 2015) describes the Xrd shader as a step function of a threshold, a light vector and a normal, with artists painting the threshold per vertex. Blender documents a Color Ramp (Constant interpolation) after Shader To RGB (EEVEE only, with stated PBR costs) as a toon shader. Unity Toon Shader documents base/1st/2nd shade colours with step and feather controls and shade colours independent of light colour. The sources do not give numeric defaults; they treat the look as an art-directed choice.
• articleWhere is AI documented in 360-film attention work (saliency prediction, script-driven object highlighting, generated 360 scenes), and what evidence supports it?
  Two documented AI connections were found. (1) Sitzmann et al. adapt existing 2D saliency predictors to 360 panoramas by multiplying with a learned equator bias and propose saliency-based cut alignment (static-panorama data, not reproduced). (2) da Silva et al. (SVR 2024, arXiv 2026, CC BY 4.0) chain Grounding DINO and Segment Anything to find the object named in a text script and apply a vignette to it in a 360 tour; their text shows frames from a university video tour, one SAM failure, and a conclusion that 'user experience' improves, but the extract shows no user study or measurement supporting that claim. No source linking generated 360 scenes to viewer-attention outcomes was found in a bounded search.
• articleWhat public head-movement and saliency datasets exist for 360-degree content, and what do they show about where viewers look (and what they cannot show about story)?
  Two open-access datasets papers were read: Sitzmann et al. (arXiv v2; 22 computer-generated static stereo panoramas, 169 users, 1980 gaze-and-head trajectories, three viewing conditions) report an equator bias and strong between-user agreement; Nasrabadi et al. (MMSys 2019; 28 videos built on a moving-object by camera-motion taxonomy, 60 participants, no audio; MIT-licensed GitHub repo with traces, questionnaires and scripts) report preliminary signs that viewers tend to look at moving targets, with exceptions, and more dispersed viewing in higher-altitude shots. Nasrabadi et al. also list other datasets with participant counts (e.g. 59 participants/5 videos; 48 subjects/20 videos; 57 subjects with eye data) and note viewing is more scattered under free viewing than when questioned afterward. These datasets describe viewing of short, mostly narrative-free content and do not test story comprehension or missed events.
• articleWhat information would a reader need before treating Wayve's GAIA-3 'predicts relative policy performance' and GAIA-4 'outcome fidelity' statements as a validated correlation with real-world driving, and which of it is missing?
  SYNTHESIS from the pages read. GAIA-3 and GAIA-4 describe validation designs (GAIA-3: simulated interventions compared with on-road experiments plus LiDAR alignment; GAIA-4: outcome, closed-loop and component fidelity on a world-on-rails replay of logged scenes) but give no population, statistic or uncertainty. A reader needs at least: the policies compared, the scenarios and their selection, the real-world metric used, the agreement statistic with intervals, who ran it, and whether the model was trained on the logged scenes it is tested on. None of these is in the page text; the only published GAIA evaluation protocol (GAIA-2 paper) tests visual and conditioning fidelity, not policy-ranking agreement. GAIA-4 states its replay keeps other road users on logged paths, which limits what an outcome agreement can show.
• articleDoes Wayve publish a technical report or paper for GAIA-3 or GAIA-4, and which GAIA versions do have one?
  In the sources inspected, Wayve's GAIA hub page lists two downloadable papers (GAIA-1, 29 Sep 2023; GAIA-2, 26 Mar 2025) and none for GAIA-3 or GAIA-4. The GAIA-3 (2 Dec 2025) and GAIA-4 (3 Aug 2026) blog pages contain no arXiv, PDF, DOI, GitHub or Hugging Face links, and an arXiv title-field search at 10:21Z found GAIA-1 and GAIA-2 but nothing titled GAIA-3 or GAIA-4. So the only evidence for GAIA-3/4 an agent can read is the company page text; the GAIA-2 technical report is the latest GAIA paper found. This is a bounded negative result, not proof that no report exists anywhere.
• articleI loaded a PC2 file in Blender's Mesh Cache modifier: which Time Mode should I use, and what do Frame Start and Frame Scale do?
  At main commit f0e5cc32 (observed 2026-10-22), Frame mode treats the evaluated value as a sample index: index = Frame Scale x scene frame - Frame Start, and ignores the file's fps and start frame (manual and RNA tooltip). Blender's own PC2 exporter manual defines 'Sampling' in frames (0.1 = 10 samples per frame), and the only tracker record about this (#41118) says a cache exported at sampling 0.1 needs Frame Scale 10 to play in real time. So for a cache with s frames per sample, Frame Scale = 1/s gives real-time playback; Time and Factor modes are not recommended for PC2 until checked (candidate 03). This is a code and document reading, not an execution.
• articleIs there a Blender bug report, commit message or regression test about the Mesh Cache modifier's PC2 Time-mode fps handling (frame = ((time / fps) - start) / sampling) or its Factor-mode scaling?
  Scoped negative result. In the sources searched on 2026-10-22 (path-filtered commit lists for six Mesh Cache files from 2013-01-21 to 2026-06-15, GitHub commit search, projects.blender.org issue and PR search, and the tests/ tree at main commit f0e5cc32) none addresses PC2 Time-mode fps handling or Factor-mode scaling. No commit message mentions fps for these files except a mechanical FPS-macro replacement (2025-08-07). No test file for the Mesh Cache modifier, PC2 or MDD was found. The nearest records are issue #41118 (PC2 sub-frames, resolved as an exporter bug in 2014, mentioning Frame Scale), #60824 (units list naming Mesh Cache frame_start) and #34607 (MDD interpolation).
• articleDo other baselines in these tables drift like DriveDreamer (version, rounding, row choice, quoted versus recomputed), and do the citing papers' own results change between arXiv versions?
  Yes for some, not all. DriveGAN has the same v1/v2 split as DriveDreamer (inside DriveDreamer's own paper) and later tables read use only the v1 pair. MagicDrive appears with at least five different values in the tables read (FID 16.20 and 16.59 in its own table; 19.1/218.1 in UniMLVG, a pair not found in MagicDrive v1/v7 or MagicDrive-V2 v4; FVD 218.12 in MagicDrive-V2; FVD 221.90 in DiVE; FID 46.18 recomputed by Delphi). Drive-WM is printed at FID 15.8 in its own text, DriveDreamer-2 and Delphi, but 15.2 in UniMLVG v1-v3 (cause not stated). Panacea 16.96/139 appears as 16.9/139.0 in DriveDreamer-2 and 16.96/139.0 in Delphi; DrivingDiffusion 15.83/332 appears rounded to 15.8 in DriveDreamer-2 and GenAD. Version drift of a citing paper's own result occurs in UniMLVG (8.8/60.1 in v1 and v2, 5.8/36.1 in v3, baselines unchanged); no change was found between the versions read for GenAD, DriveDreamer-2, Delphi (v1 vs v3), MagicDrive-V2 (v1 vs v4) or MagicDrive (v1 vs v7 for its MagicDriv...
• articleBefore an agent uses a free 3D, texture, HDRI or sound asset in a production or dataset, what per-asset licence evidence can it record, which libraries expose a machine-readable licence field, and which must be captured from pages?
  A synthesis proposal built from the observed library structures. Sketchfab exposes a licence object per model; the Poly Haven and ambientCG API responses inspected carry no licence field, so their licence evidence is a site-wide page; Freesound extends each sound's licence to its metadata and keeps uploader AI preferences per user; Sonniss and Sketchfab say the version in force at download governs; BBC limits uses by purpose. The record therefore needs the asset id, licence label and URL, the licence and site-terms texts with hashes and dates, the version rule, AI-related flags, and per-component status, all with access time. This is a proposal, not a tested procedure and not a legal conclusion.
• articleWhat exactly did the WHAM paper measure for persistency, with what values, sample sizes, labelers and limits?
  Nature Table 1 (observed 2026-10-22) gives the share of 1.6B-WHAM generations labelled as persisting an edited-in element: with one conditioning frame, Powercell 58%, character 45%, Vertical Jumppad 58%; with five frames, 86%, 85%, 98%. The protocol is small and narrow: 600 videos (480 for Powercell/character, 120 for Jumppad), no-op actions, 10 sampled videos per edited input, each judged only over the first ten generated frames (1 s) by seven of the authors, blinded to input length, with 90% agreement on the 300 double-labelled videos (stricter label kept). The 10-frame condition appears only as an Extended Data figure (values not extracted), and the text says persistency rose from 1 to 5 frames but not from 5 to 10.
• articleDo Maya, Unreal Engine and Houdini document ML-based deformation or simulation tools, and do they replace or complement the traditional rig/solver?
  Yes for deformation, and in every page read the ML model is trained on the output of the existing rig or simulation and complements it. Maya's ML Deformer approximates a complex deformer for posing and playback or crowds and keeps the original deformer for final rendering; Unreal's ML Deformer approximates deformations at runtime from data made in Maya or Houdini and recommends Neural Morph for flesh-like and Nearest Neighbor for cloth, and Epic says Fortnite used it for muscles and cloth; Houdini (docs labelled 22.0) has ML Deform, an ML Deformer recipe that approximates an expensive flesh simulation, and an ML Volume Upres recipe for pyro. No vendor page read documents ML replacing a cloth or fluid solver; Houdini ML cloth was not found in pages read.
• articleDo learned physics simulators stay stable over long rollouts and generalise beyond their training data, and what do the sources say breaks?
  Both papers report plausible rollouts far longer than the training trajectories (GNS: thousands of steps; MGN: 40,000 steps from a model trained on 400-step trajectories) after training on one-step targets, and GNS credits input-noise injection for this. Generalisation claims are specific: GNS to about 10x more particles and unseen shapes, MGN to a 20k-node windsock and wider angle/speed ranges with RMSE 11.5 to 12.4/13.1. Failures are also stated: GNS rigid block deformation in vigorous shaking and a stuck-goop bad seed; MGN cloth decoherence after about 50 steps; a baseline GNS without mesh-space features is unstable on cloth. A later PyTorch re-implementation README warns accuracy may degrade over time and extrapolation in time should be avoided.
• articleWhat is a 'learned physics simulator' such as DeepMind's GNS or MeshGraphNets, what does it take as input and predict, and what is it trained on?
  In the two DeepMind papers a learned simulator is a graph neural network trained, per scenario family, on trajectories produced by a conventional solver (MPM, SPH, PBD, ArcSim, COMSOL, SU2). It reads the current state (particles or mesh nodes with a short velocity history), predicts per-node acceleration, integrates it with an Euler step, and is applied repeatedly to roll out a trajectory. It is a fast surrogate of the solver it was trained on, not a model that discovers physics; the training supervision is one-step prediction.
• articleWhat do the primary papers leave unshown for collaborative previs or multiplayer games with a generated shared world, according to their own stated limitations and evaluation protocols?
  Each paper states a limit that matters for collaborative previs or multiplayer play: Solaris and MIRA lose persistent state beyond the context; Gamma-World may accumulate inconsistencies in long rollouts; MetaWorld does not model physical contact; WorldCast does not model hits or network faults; MultiWorld has not been trained at scale. Beyond these, in the eight papers read no paper reports art-direction or director controls, a multi-human study, or output-use rights for generated clips. The last three are absences in the passages read, not demonstrated impossibilities.
• articleWhat scale (players or agents, rollout length, resolution, hardware, speed) do the primary papers actually demonstrate for a shared multi-agent world, and which numbers are measured versus shown only in figures?
  Measured consistency is shown for two players in most papers (Solaris, ShareVerse, MetaWorld two views) and up to four players only qualitatively (Gamma-World zero-shot) or with a game-state probe and real-time claim (MIRA, 4 players, 20 FPS on one B200). WorldCast times 2 to 16 players per-client (FPS) but scores consistency in rounds with three to five clients. MASS reports 1,024 simulated players, but these are entities in a grid game with 128 by 128 renders, not 1,024 rendered video players.
• articleHow do the audio-capable world-model papers measure that sound stays in sync with actions, camera motion and events, and what do they leave unmeasured?
  Three different measurement designs exist. AVWM compares the audio-energy vs motion-energy cross-correlation lag and peak of generated audio with that of ground-truth audio against the true visual trajectory (simulator data, 0.15 s steps); HelixBench (1,015 clips) scores DeSync on an Event subset and a Grounding-DINO-based three-region stereo score for camera-aligned panning; MaineCoon reports an unexplained "AV-Al" column where it is behind Ovi. Each protocol scores windows or short rollouts; none of the texts read measures the delay from a user action to the sound onset.
• articleWhat do the 4D reconstruction and generation papers say about relighting, editing, topology and render cost, i.e. the limits that matter for VFX, volumetric capture or virtual production?
  Across nine primary papers (HTML bodies, before reference lists) no text contains the words relight, albedo, BRDF or virtual production; lighting appears only as scene variation to be fitted (Real-time 4DGS) or as an uncharacterised factor (DyCheck). Gaussian colour is stored as spherical-harmonic coefficients, so appearance is baked and view-dependent; relighting is therefore not a documented capability of these systems. Editing is demonstrated for rig-captured persistent Gaussians (edit propagation, object insertion, camera attachment) and described as naive transformation in 4D-GS. Topology: persistent Gaussians give point correspondence over time, DG4D gives a per-frame obj mesh sequence, and no read source reports a stable mesh topology. Render speed is high for small scenes on a consumer GPU (author-reported tens to hundreds of FPS) but is tied to Gaussian count, resolution and deformation network size; clips are seconds long and scale to large scenes is a stated limit. Produc...
• articleWhen one model is scored under several action-following protocols (Oasis and Matrix-Game across MineWorld, Matrix-Game, GameCraft and Yume papers), do the scores agree, and are the protocols comparable?
  They do not agree. Oasis keyboard accuracy is 0.77 in Matrix-Game 1.0, 0.73 in Matrix-Game 2.0, and F1 0.41 (precision 0.49, recall 0.44) in MineWorld's own table. Matrix-Game 1.0 scores 0.95 keyboard/0.95 mouse in its own paper, yet in GameCraft's tests (150 mixed images, 12 actions) it has the highest translation error and the second-highest rotation error of five methods (RPE trans 0.18, rot 0.35 versus 0.08 and 0.20 for GameCraft) and the lowest user-study action-accuracy rank (1.63 of 5), and in Yume-Bench its human-rated instruction following is 0.271 against Yume's 0.657. The sources differ in test set, domain, estimator and rater, so these are four separate protocols, not four measurements of one quantity. The differences are consistent with domain shift (Yume says Matrix-Game struggles to generalise to real-world scenes) but no source isolates the cause.

More