3D and worlds

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

• articleWhat did DeepMind report when SIMA 2 agents acted in Genie 3-generated worlds: inputs, outputs, constraints, measured results?
  The SIMA 2 paper (DeepMind, Dec 2025) runs a Gemini Flash-Lite-based agent on 720p frames with keyboard-and-mouse actions in Genie 3 worlds. Genie 3 environments are text- or frame-conditioned, split into urban (train) and natural (held-out) navigation tasks. Qualitatively the agent navigates to points of interest; for self-improvement the authors report gains on most train tasks, often 25 points or more, and on most held-out natural tasks (Fig. 17; values read from figure not extracted). The Genie 3 blog (Aug 2025) is an earlier company statement of a SIMA run with no figures. Task count, number of episodes and variance are not given in the text I extracted.
• articleWhat does the public MPEG V-PCC/V3C reference software (mpeg-pcc-tmc2 release-v25.0) say about its role, build dependencies, how occupancy, geometry and attribute video are encoded, and under what licence?
  MPEG's V-PCC reference software (TMC2, release-v25.0) is an experimental reference, not an efficient implementation: it codes a point cloud as separate occupancy, geometry and attribute 2D video streams using external video codecs selected by codecId (JM, HM, VTM, SHM, or FFmpeg on a branch), with stacked configuration files where later files override earlier ones; its test example expects an external MPEG dataset path. The code is under an ISO/IEC BSD-style licence that expressly grants no patent rights, so patent clearance is a separate question.
• articleHow did DreamGen use video world models to create synthetic training data for robot policies, and what did it measure?
  Partially verified: the curator re-read none of the cited sources; S10 rests on the producer's pinned reading. DreamGen (NVIDIA et al., arXiv v2, June 2025) fine-tunes an image-to-video model on a target robot, rolls it out from an initial frame plus instruction, recovers pseudo-actions with an inverse-dynamics model or latent actions, and trains a policy on these 'neural trajectories'. Reported results: on a humanoid with GR00T N1, 14 novel behaviours 11.2% -> 43.2% and 13 novel-environment tasks 0.0% -> 28.5% (policy trained only on neural trajectories plus a single-environment pick-and-place baseline); RoboCasa solely-synthetic training reaches 20.6% average over 24 tasks; policy performance rises log-linearly with number of neural trajectories. Costs and limits: 240k RoboCasa samples took 54 hours on 1500 L40 GPUs; initial frames are provided manually; the benchmark's automatic evaluator can hallucinate.
• articleWhat do the NVIDIA Cosmos paper and docs actually document for synthetic data generation and agent evaluation, versus only propose?
  Partially verified: the curator re-read none of the cited sources; S5, S7, S8, S9 rest on the producer's pinned reading. The Cosmos paper (v3) lists four uses of a world foundation model (policy evaluation, policy initialization/training, planning, synthetic data) and says it reports no empirical results for them. What it does measure are model-level checks: 3D consistency, physics alignment and post-training results for camera control, robotic manipulation (action-conditioned and instruction-conditioned video prediction) and driving. The docs (observed 2026-10-22) describe workflows with inputs/outputs: Cosmos-Transfer2.5 for two augmentation workflows (Simulations to Photorealism; Scale World State Diversity) with computed control maps; Predict2.5 action-conditioned post-training on Bridge end-effector poses; Reason1 best-of-N rejection sampling with a score = fraction of artifact-free evaluations. The docs show example videos and procedures but no downstream policy or accuracy re...
• articleBefore using a generated world (Marble splat or mesh, HunyuanWorld layers, WonderWorld splat) for previs or as an LED-volume background, what should an agent check, what does the one documented LED-volume use report, and what is unknown?
  Partially verified: the curator re-read none of the cited sources; S25, S20, S9, S15, S16, S21, S5 rest on the producer's pinned reading. The only documented LED-volume use found is a vendor case study (Volinga, Feb 2026): a music video used Marble worlds shown as Gaussian splats in Unreal via the Volinga plugin, with stYpe camera tracking and set extension; the vendor and its customer report time saved and on-set switching, with no data on colour, resolution, latency or artifacts. World Labs' docs add plugin notes for nDisplay and depth of field. WonderWorld's authors say its own results have low density and floaters and suggest it as a coarse prototype to refine. The checks below are proposals synthesised from these statements, not tested procedures.
• articleWhich benchmark checks whether a video or world-generation model follows a camera path and keeps a scene consistent, and does it measure physical laws?
  Partially verified: the curator re-read none of the cited sources; S3, S10, S11 rest on the producer's pinned reading. WorldScore (Stanford; arXiv v2 Nov 2025) evaluates 'next scene' generation: given a scene image, a next-scene prompt and a camera layout (a trajectory plus one of 8 text camera moves), it scores controllability (camera, object, content), quality (3D, style, photometric consistency, subjective quality) and dynamics (motion accuracy, magnitude, smoothness). It tested 20 models, including 13 video models; video models had weak camera control (best video 40.22 on camera controllability) and 3D scene generators beat them on static worlds. It does not define a physical-law metric, so it informs camera and consistency questions only.
• articleDo the Objaverse-trained 4D generators (L4GM, SV4D, SV4D 2.0, Diffusion4D) say whether their training data overlaps the Consistent4D benchmark objects, and what should an agent assume when comparing their scores?
  Partially verified: the curator re-read none of the cited sources; S14, S13, S9, S5, S15, S16, S8 rest on the producer's pinned reading. The Consistent4D author states that six of the seven benchmark objects are in Objaverse and recommends filtering them from training. A set intersection I ran (synthesis, not executed code from the repositories) finds 6 of the 7 benchmark uids in a 37,789-ID list of animated Objaverse-1.0 objects published in the Diffusion4D repository, matching that statement. SV4D states it excluded the benchmark objects; L4GM v1 states that it renders all animated Objaverse 1.0 objects and evaluates on the benchmark, and the passages searched contain no exclusion statement; SV4D 2.0 v3 contains no such statement in the strings searched. Overlap is therefore an open question for L4GM, and possibly SV4D 2.0, on this benchmark, not an established fact: L4GM filters videos by motion, and the L4GM training list is not in its repository.
• articleWhat exactly does HunyuanWorld 1.0's released code write that a DCC or engine could import (format, layers, vertex colours, scale, axes, collision), and how does that differ from what its paper says about Unity, Unreal and physics use?
  Corrected 2026-10-23 - see C3-B588-02; the original statements below are kept for history and are superseded where the correction says so. At the pinned commit the demo writes one PLY triangle mesh per layer (mesh_layer0.ply, ...; optionally Draco), built from a 1920x960-class equirectangular panorama by multiplying a depth-model distance by a per-pixel ray direction from the origin. Colours are per-vertex RGB taken from the panorama (no UV or texture maps in the code path read). Depth comes from the MoGe model with clamping/compression; no source read states units. No collider or engine-specific export is written. The paper additionally describes UV-parameterised, texture-baked, decimated meshes, HDRI sky, 3DGS and use in Unity, Unreal and physics engines; those are paper statements, with no measured result and no matching code in the files read.
• articleWhat do World Labs' docs state about Marble exports entering a DCC, engine or virtual-production pipeline: file formats, units and scale, coordinate convention, baked vs dynamic lighting, and collision?
  Corrected 2026-10-23 - see C3-B598-02; the original statements below are kept for history and are superseded where the correction says so. Marble exports splats (PLY, SPZ) and GLB meshes. Docs state: splats need third-party plugins in Unreal, Unity, Blender and Houdini; collider meshes (100-200k triangles) are for coarse physics, not visual rendering; units are 'arbitrary model units' in API exports, with a metric_scale_factor and ground_plane_offset supplied for conversion; and the coordinate convention is documented inconsistently across pages and dates (OpenCV-style 'raw' for SPZ, OpenGL default from Dec 2025, an OpenGL/OpenCV option from Jan 2026). No inspected page says whether lighting is baked into splat colours; plugin pages describe relighting and shadows as plugin-dependent.
• articleIf I only have a text prompt, what do MAV3D and 4D-fy actually show, how were they evaluated, which one can I run, and how good is the motion?
  Corrected 2026-10-23 - see C3-B588-07; the original statements below are kept for history and are superseded where the correction says so. Partially verified: the curator re-read none of the cited sources; S1, S3, S4, S19, S8, S16, S9, S5 rest on the producer's pinned reading. Two papers cover text-to-4D with score distillation. MAV3D (Meta, January 2023) optimises a six-plane dynamic NeRF against a text-to-video model in three stages (about 6.5 h on 8 A100) and shows an R-Precision of 83.7 against adapted internal baselines of 6.2-10.6; no code is available. 4D-fy (arXiv v2, May 2024) blends 3D-aware, image and video diffusion priors, takes about 23 h on one 80 GB A100, and ships a repository with run instructions. In 4D-fy's own user study (26 evaluators, 28 prompts) users prefer 4D-fy for appearance (92%) and structure (89%) but only 41% prefer it for motion, and the motion difference is the one comparison that is not statistically significant. 4D-fy says motion is limited to sim...
• articleHow many frames, views and what resolution does each text/image/video-to-4D system produce, and what GPU time and memory do the authors report?
  Corrected 2026-10-23 - see C3-B588-07; the original statements below are kept for history and are superseded where the correction says so. Partially verified: the curator re-read none of the cited sources; S1, S11, S3, S4, S13, S14, S15, S16, S17, S5, S6, S8 rest on the producer's pinned reading. Reported per-asset cost spans about four to five orders of magnitude, all author-measured on different hardware. Seconds: L4GM (0.3 s per 16-frame pass plus about 2 s for views, RTX 4080 Super). Minutes: DreamGaussian4D (6.5-10 min, 80 GB A100), Diffusion4D (8 min, paper), SV4D (15-20 min fitting) and SV4D 2.0 (1 min generation + 21 min fitting, up to 30 GB, H100). Hours: Consistent4D (about 2-2.5 h, V100), MAV3D (about 6.5 h on 8 A100) and 4D-fy (about 23 h on one 80 GB A100). Clips are short: 12-32 frames per pass depending on the system (12, 14, 16, 21, 24, 32), time axes are continuous for the NeRF/neural-field systems, and no passage read states output duration in seconds for text-to-4...
• articleWhich of the driving and robot world models can an agent actually download and run, on what hardware, and what does each component's text say?
  Corrected 2026-10-23 - see C3-B598-04; the original statements below are kept for history and are superseded where the correction says so. Open and runnable: Vista (weights on Hugging Face, 32 GB VRAM for sampling), Cosmos-Predict1 driving multi-view checkpoints (three types in the downloader) and Cosmos-Predict2.5 driving multiview and robot action-conditioned models (README/table listed), and the 1X challenge GENIE baselines. Code only: DriveDreamer (research code; no weights named in README/DOCS). Not found: GAIA-1/2/3/4 (no repository, release or access statement found) and 1XWM (no release stated). Hardware is large for the Cosmos multiview models (8 GPUs of 80 GB for Predict2.5 multiview inference). Rights differ per component and there is one same-work tension: Vista's card says Apache-2.0 while its paper says it was initialised from SVD, whose checkpoint the authors call non-commercial. Cosmos licence clauses are covered by C3-B019-W6 and are not repeated.
• articleIf I have one video (or one image) of an object and no rig, which video-to-4D generation systems can produce an animated 3D object, what do they output, and what do they cost to run?
  Corrected 2026-10-23 - see C3-B588-09; the original statements below are kept for history and are superseded where the correction says so. Four released video-to-4D lines were read at pinned versions. Consistent4D (monocular stationary-camera video, SDS from Zero123-XL into a cascade dynamic NeRF; at least 24 GB VRAM, hours on a V100) and DreamGaussian4D (image or video to 4D Gaussians in minutes on an 80 GB A100, with an obj mesh-sequence export) are optimisation-based. SV4D and SV4D 2.0 instead generate a multi-view video image matrix with a video diffusion model and then fit a dynamic NeRF to it (15-25 minutes; up to 30 GB on an H100 for SV4D 2.0). All are object-centric; SV4D 2.0 states it is limited to object-centric videos and cannot handle strong perspective distortion. Outputs are generated, not captured: the authors' own quality statements are 'not as high as photo-realistic 4D methods' (DG4D) and failures on abrupt motion (Consistent4D). Weights and rights are in candidate...
• articleWhat representation does each of Marble, HunyuanWorld 1.0, WonderWorld and WonderJourney actually output (splats, meshes, panoramas, depth layers, point clouds, video), and is each a paper result, a demo or a released capability?
  Corrected 2026-10-23 - see C3-B588-01 and C3-B598-01; the original statements below are kept for history and are superseded where the correction says so. The four systems output different things. Marble (World Labs) is a hosted product whose docs list Gaussian splats (SPZ/PLY), triangle meshes (GLB), a 360 panorama and rendered video. HunyuanWorld 1.0 builds layered meshes from a generated panorama plus depth; its paper also describes a 3DGS and an HDRI-sky option, but the released demo code at the pinned commit writes only per-layer vertex-coloured PLY meshes (optional Draco). WonderWorld produces Fast Gaussian Surfels (saved as .ply and .splat by its code); WonderJourney produces a sequence of coloured point clouds that its code renders to video. RTFM, also from World Labs, deliberately has no explicit 3D representation. Labels differ by component, see systems[].
• articleWhich open image-to-3D and text-to-3D asset generators have released code and weights, and what file or representation does each one produce?
  Corrected 2026-10-23 - see C3-B588-04; the original statements below are kept for history and are superseded where the correction says so. Eight open releases were read at pinned commits. Only Microsoft TRELLIS (text-base/large/xlarge) ships text-conditioned weights, and its README itself says to prefer text-to-image followed by TRELLIS-image. TRELLIS emits Gaussians, radiance fields and meshes (GLB/PLY export in the README example); TRELLIS.2 emits a mesh with PBR maps (GLB); Hunyuan3D 2.x emits a trimesh shape then a separate texturing stage (2.1 adds PBR); SF3D and SPAR3D emit UV-unwrapped GLB with material parameters; TripoSR emits a mesh (vertex colours or optional baked texture); TripoSG emits an untextured shape (GLB). Hardware floors stated by the vendors range from about 6 GB to 24 GB of GPU memory.
• articleWhat do primary sources say, measured or claimed, about topology, UVs, polygon budgets and PBR correctness of generated assets for games, previs or VFX?
  Corrected 2026-10-23 - see C3-B588-05; the original statements below are kept for history and are superseded where the correction says so. No inspected source measures retopology quality (edge flow), UV distortion or engine polygon-budget compliance, and none tests PBR values against ground-truth materials of a held-out asset. What exists: SF3D reports triangle counts for one mesh per method (SF3D 27.4K vs TripoSR 32.1K, OpenLRM 662K) and a box-projection UV unwrap in 0.5 s, with remesher options (triangle, quad split to triangles); TRELLIS.2 reports PSNR/LPIPS on rendered PBR maps for reconstruction, offers GLB export with remesh and a decimation target (README example 1,000,000 faces, 4096 texture), and states small holes occur; SF3D and SPAR3D state material estimation is homogeneous or sometimes suboptimal; TRELLIS (first release) states baked-in lighting and no PBR; Hunyuan 2.0 describes a low-polygon stylization module in its hosted Studio. Production claims are the vendors'.
• articleWhat licences and access terms apply to Marble, HunyuanWorld 1.0, WonderWorld and WonderJourney (code, weights, outputs), and what hardware does running each require?
  Corrected 2026-10-23 - see C3-B588-03; the original statements below are kept for history and are superseded where the correction says so. Four different regimes. Marble is a hosted service under World Labs Terms of Service (last updated 2026-01-21): Free-account output is retained by World Labs with a personal non-commercial licence to the user; paid-account output is owned by the user including commercial purposes, with a training licence over user content that paid users may opt out of. HunyuanWorld 1.0 is under a Tencent community licence that is expressly limited to a Territory excluding the EU, UK and South Korea, includes a 1M monthly-active-user trigger and a bar on using outputs to improve other AI models. WonderJourney code is MIT. WonderWorld has no licence file at the pinned commit. Nothing here is legal clearance.
• articleI have ARKit or MediaPipe blendshape values and need FACS action units: what published mappings exist, how were they made, and which should I trust first?
  Three mappings exist and are made differently: a practitioner ARKit table (Ozel), an expert-consensus MediaPipe-52 table (CNR clinical pre-proof, no validation against real facial behaviour data) and a learned AU-to-blendshape model (ICAT-EGVE 2024 poster). None was validated against manual FACS coding of the same footage, so any table is a hypothesis to check on your own clips; keep one-to-many cases (eyeSquint AU7 with AU44; cheekPuff AU33/AU34; jawOpen AU26/AU27) as multi-label.
• articleHow do Blender's Mesh Cache modifier time modes (Frame, Time, Factor) map to PC2 and MDD frame positions, and do PC2 and MDD agree in seconds mode?
  In Scene play mode Frame time is the scene frame scaled and offset, used directly as cache index by both readers. In Time mode the modifier passes ctime/fps (seconds) and the PC2 reader divides by fps again before applying start/sampling, while MDD compares seconds to stored times; by reading, PC2 Time mode is therefore inconsistent with MDD (code inference, not run). Curator addition: Scene-mode Factor also uses ctime/fps, so Factor time = frame_scale*(ctime/fps) - frame_start, so the clamped factor reaches 1.0 after (1 + frame_start)/frame_scale seconds of scene time, not at the last cache frame (code inference, not run; amendment CA-AM-0036).
• articleWhich glTF material inputs have no OpenPBR counterpart in the texts read, and which OpenPBR features have no glTF counterpart?
  OpenPBR (1.1.1, still current) and glTF materials don't line up. OpenPBR has no terms for glTF's occlusion texture, alpha cutoff or unlit mode. glTF has no counterpart for OpenPBR's diffuse roughness, subsurface scattering controls or coat darkening. OpenPBR also lets renderers approximate costly features, so some 'loss' is a renderer choice. MaterialX publishes translation graphs that chain OpenPBR -> Standard Surface -> glTF PBR (route inferred from matching input/output names; neither .mtlx file describes the chain in prose; read, not run; amendment CA-AM-0053). The first step carries diffuse roughness, subsurface, coat IOR and coat darkening into Standard Surface, and the second step has no inputs for them, so they are dropped on export to glTF. Coat colour is folded into an attenuation term rather than dropped.
• articleWhat scores do named systems (Sora 2, Veo 3.1, Gen-3, Hailuo, Cosmos, SVD, LTX-Video and others) get under each geometric-consistency metric, and which of those numbers can legitimately be compared?
  Corrected 2026-10-23 - see C3-B588-06; the original statements below are kept for history and are superseded where the correction says so. GeCo-Eval (80 static-scene prompts, 4 videos each, with motion columns) is the only inspected benchmark that scores Sora 2 and Veo 3.1; it puts Sora 2 and Veo 3.1 close together (object-centric fused 0.069 vs 0.090, indoor 0.169 vs 0.233, outdoor 0.265 vs 0.217; lower is better) with Veo moving more. WorldScore gives Gen-3 68.31 and Hailuo 67.18 on its 0-100 reprojection axis; SGC scores Cosmos 0.0722 and Latte 0.3226; the MEt3R paper has SVD at 0.032. Metrics, scales, prompts, clip lengths and motion differ, so the numbers form a table of separate experiments, not a ranking; the only visible overlap is model family names.
• articlePer component (tokens, raw video, weights, code) do the 1X cards, the GitHub README/LICENSE and the 1X blog agree on terms, and which uses does each text actually address?
  Corrected 2026-10-23 - see C3-B588-08; the original statements below are kept for history and are superseded where the correction says so. For each component every text read gives the same label: tokenised data Apache 2.0 (HF card, GitHub README schema block, 1X blog), raw video CC-BY-NC-SA 4.0 (raw card front matter, tokenised card prose, 1X blog), GENIE weights apache-2.0 (card front matter/API), code Apache-2.0 (LICENSE text). Different labels on different components are not a conflict. No text addresses hosting, redistribution, modification, training, commercial or output use except by the label itself; stated uses are purposes (experimenting with robotics data, challenge participation) and a contest rule. No rights conflict is filed.
• articleProperty by property, what do the OpenPBR 1.1.1 specification and the Khronos glTF material READMEs each define for the properties a glTF to OpenPBR conversion must carry?
  OpenPBR v1.1.1 on main (f8d6d947, merged 2026-04-17) contains no glTF text and no retroreflection, occlusion or alpha-cutoff terms, so no source read states a glTF-to-OpenPBR mapping. The glTF retroreflection extension (Release Candidate) says it matches OpenPBR's specular_retroreflectivity, a parameter merged on 2026-02-23 into OpenPBR's dev_1.2 branch, not v1.1.1, which explains the producer-flagged contradiction. glTF dispersion stores 20/V_d (stated as OpenPBR's transform) and emissive strength is a unitless multiplier, while OpenPBR emission_luminance is in nits.
• articleWhat do World Labs' current docs say about Marble units and default coordinate convention, and does the 'arbitrary model units' / 'OpenGL default from Dec 2025' wording hold on the cited pages?
  Correction to C3-B012-W4-05 (nature: changed or unsupported on the cited pages). Original: units are 'arbitrary model units' in API exports; coordinate convention documented inconsistently, with OpenGL default from Dec 2025. Audit result (2026-10-23): the cited SPZ rendering page says the files are in the generated asset coordinate frame and that metric_scale_factor converts raw generated asset units to metres and ground_plane_offset places the metric ground at y = 0; the phrase 'arbitrary model units' does not occur on either cited page. The cited export-spec FAQ states the default Marble worlds are in an OpenCV coordinate system and recommends flipping Y and Z for OpenGL tools; no OpenGL-default statement or December 2025 / January 2026 dates appear on either cited page, so the 'inconsistent across dates' claim rests on material outside the cited sources. Agents should apply the scale metadata (metres) and the OpenCV-to-OpenGL flip as documented now, and re-check if a changelog is...
• articleWhat does World Labs' Marble export-spec page actually list (panorama size, mesh tiers and their limits), and is video part of that list?
  Correction to C3-B012-W4-01 (nature: missing limitation; one item not on the cited page). Original: the docs list splats, GLB meshes, a 360 panorama and rendered video. Audit result: the cited export-spec page (checked 2026-10-23) lists a prompt image, a 2560x1280 equirectangular PNG panorama, SPZ and PLY splats at about 2M and 500k, a coarse collider GLB (100-200k triangles) and a high-quality mesh tier. The high-quality tier is two GLBs (about 600k triangles with texture, about 1M with vertex colours), takes up to an hour, is rate limited to 4 generation requests per hour per user and only for worlds the user owns; the original omitted these limits. The export-spec page does not mention video. Video appears in the docs index as a studio Record tool for flythrough videos, a different feature from export files, so 'rendered video' should be attributed to Record, not the export specs. Version/date: docs as fetched 2026-10-23.
• articleExactly how does DreamGaussian4D describe its generation quality relative to other 4D methods?
  Correction (overstated quote). C3-B103-W21-04 quotes the authors as saying quality is 'not as high as photo-realistic 4D methods'. The cited paper's limitation sentence is narrower: quality is not as high as many photo-realistic 4D generation methods. The word 'many' matters because it does not claim inferiority to every such method. Version/date. arXiv 2312.17142v3 fetched 2026-10-23.
• articleWhat do the cited Diffusion4D and 4D-fy papers report about generation time, and where do the MAV3D 6.5-hour and Diffusion4D 8-minute figures come from?
  Correction (unsupported by the cited source). C3-B168-W33-03 and W33-04 give Diffusion4D an '8 minutes' text-to-4D time and MAV3D about 6.5 hours on 8 A100. In the cited arXiv pages (Diffusion4D 2405.16645v1, MAV3D 2301.11280v1) the text extracted says only that Diffusion4D generates 4D assets 'within just several minutes' and contains no A100 or hours figure for MAV3D. The 4D-fy time is supported: the three optimisation stages take roughly 2, 2 and 19 hours (23 hours total) on an A100 with about 80 GB of VRAM. Treat '8 minutes' and '6.5 h on 8 A100' as unverified (tables or later versions may carry them). Version/date. Fetched 2026-10-23.
• articleWhere are the TRELLIS.2 small-hole limitation and the SF3D, SPAR3D and TRELLIS v1 material or baked-lighting statements actually stated?
  Correction (unsupported by the cited source). C3-B080-W16-07 attributes several limits to the cited READMEs: TRELLIS.2 'states small holes occur', TRELLIS first release 'states baked-in lighting and no PBR', and SF3D/SPAR3D 'state material estimation is homogeneous or sometimes suboptimal', plus SF3D triangle counts and a 0.5 s UV unwrap. In the pinned READMEs fetched 2026-10-23 none of these statements appears. The TRELLIS.2 small-holes statement is real but sits in the Hugging Face model card, under Known Limitations. The SF3D/SPAR3D and TRELLIS v1 statements and the SF3D numbers were not found at the cited sources (they may be in papers or HF cards not cited). Cite the card for the holes statement and treat the rest as unverified until a source is added. Version/date. Fetched 2026-10-23.
• articleWhat GPU memory do the open image-to-3D releases really need, and does it stop at 24 GB?
  Correction (overstated range). C3-B080-W16-01 says hardware floors stated by the vendors range from about 6 GB to 24 GB. The pinned Hunyuan3D-2.1 README states 10 GB for shape generation, 21 GB for texture generation and 29 GB for shape plus texture in total; TRELLIS.2 states at least 24 GB. So the upper end of the stated range is 29 GB for Hunyuan3D-2.1's full pipeline, not 24 GB. Hunyuan3D-2 (earlier) states 6 GB for shape and 16 GB total. Version/date. Pinned commits fetched 2026-10-23.
• articleDoes a paid Marble plan carry commercial rights, or only the Pro plan?
  Correction (missing limitation). C3-B012-W4-03 states that paid-account output is owned by the user including commercial purposes, citing the account and billing page. The World Labs Terms of Service (last updated January 21, 2026) do say Paid Account Users own Outputs and may exploit them for any purpose including Commercial Purposes. However, the cited plan page lists 'Includes commercial rights to generated worlds' only under the Pro plan, not under Standard (also a paid plan). The two pages therefore read differently for Standard users, and the summary did not note it. Treat the Standard-plan position as unresolved and ask World Labs before commercial use. Limitations. The ToS may define Paid Account broadly enough to include Standard; the plan page may be a marketing summary. Version/date. Both fetched 2026-10-23.
• articleWhat resolution does HunyuanWorld 1.0's world_composer use by default, and is it 1920x960?
  Correction (changed or unsupported figure). C3-B012-W4-06 says the demo builds meshes from a '1920x960-class' equirectangular panorama. In the pinned world_composer.py the default working resolution is 3840 x 1920 (width, height), the foreground and background mesh resolution caps are 3840, and only the sky mesh cap is 1920. Input panoramas of another size are resized to the configured resolution. The 1920x960 figure is therefore not what the cited file defines, though the 2:1 aspect is right. Version/date. Pinned commit 57fa9f3 fetched 2026-10-23.
• articleWhich Marble exports are documented at the cited pages, and what does HunyuanWorld 1.0's pinned README say it outputs (is video or 3DGS/HDRI documented there)?
  Correction (unsupported by the cited source). C3-B012-W4-01 says Marble's cited docs list rendered video as an export and that the HunyuanWorld paper describes 3DGS and an HDRI sky. Re-fetch 2026-10-23: the cited export specs page lists a 360 panorama (equirectangular 2560 x 1280 png), SPZ/PLY splats and GLB meshes, with no video file among them; video appears only in the plan page, where Pro adds 'enhancing the quality of generated video outputs'. The pinned HunyuanWorld README describes panoramic world proxies and a layered mesh with mesh export; it does not mention 3DGS or an HDRI sky (the paper, which was not among the cited URLs, may). Use the narrower statement unless a cited page is added. Limitations. Video generation is a documented Marble feature; only a video export format is unsupported at the cited page. Version/date. Fetched 2026-10-23.
• articleAt Alembic 2d54c7c2 the Ogawa core is always built while HDF5 is an option defaulting OFF; the library tree has no skeleton/skin module; archives can carry only an optional DCC FPS hint (_ai_DCC_FPS); LICENSE.txt declares BSD 3-Clause using a non-standard SPDX spelling, which plausibly explains...
  At Alembic 2d54c7c2 the Ogawa core is always built while HDF5 is an option defaulting OFF; the library tree has no skeleton/skin module; archives can carry only an optional DCC FPS hint (_ai_DCC_FPS); LICENSE.txt declares BSD 3-Clause using a non-standard SPDX spelling, which plausibly explains GitHub API NOASSERTION (curator observation).
• articleWorldbuilding in service of story, and using concept-art pushes to feed writing (Sanderson, Dos Santos, Stewart panel)
  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. DSNX 2025 panel (published 2026-01-23). Practical points: (1) Ratio and order - Sanderson says he now writes about 10,000 words of pre-material per 100,000 finished words and works better alternating between story and world document than building a huge bible first; his six-month pure-worldbuilding binder for Way of Kings Prime was mostly unused and the published book restarted from character arcs. (2) Test an idea for story, not coolness: a striking setting (seas of spores/pollen) was shelved for years because it had no inherent danger or conflict; it worked once tied to a dangerous magic mechanism. (3) Concept-art pushes - a group brainstorm over a fixed period, often...
• articleTwo new permissively licensed Tencent releases for creators: Hy-Nomas (Apache-2.0 interactive world video with camera control) and AuK (MIT speech generation/editing with zero-shot cloning)
  Hy-Nomas (card created 21 Oct 2026, Apache-2.0 - notably NOT the territory-restricted Tencent community licence used for HY-World 2.0, C3-B399-01): a segment-autoregressive video "world" model - each new segment is generated from text instructions and optional camera trajectories while a persistent KV cache carries context, enabling multi-turn, free-form interaction (introduce subjects, actions, style or environment changes) and camera navigation; 8-step distilled DiT with a PixelUMM decoder; default output 720x1280 at 24 fps (480x832 option); standard inference uses 8-way sequence parallelism (eight GPUs), with an INT8/FP8 single-GPU configuration that keeps a bounded history; ships Hy-Nomas-Bench (280 test cases) and Docker/offline JSON request format with timed segments. AuK (open-sourced 9 Sep 2026, MIT; 1.5B parameters; AuK-Flash 4-step variant): one natural-language instruction interface for zero-shot and instruction TTS, content/acoustic and paralinguistic speech editing, spe...
• articleDecart: priced, API-accessible real-time models - Oasis 3 Preview (text-prompted driving world model) and Lucy 2.5 live video editing at about $0.02 per second
  Decart's API platform (docs fetched 2026-10-23) is one of few places where a world model is a paid, documented API rather than a demo. Oasis 3 Preview (released June 2026, R3): set a scene with a text prompt, then send chunks of exactly four [throttle, steering] actions in [-1, 1] and receive four generated frames for each of three camera streams (left_forward, front, right_forward) over a stateful gRPC session; a new prompt resets the world context; Python SDK decart-oasis; a browser playable preview exists; the docs pitch it as a learned driving simulator for RL, with a PPO Colab (R1). It is driving-specific (vehicle controls, three car-mounted views), not a general walkable world. Pricing (R2): pay-as-you-go, no subscription; Oasis 3 Preview $0.02 per second (a 60 s session = $1.20); Lucy 2.5 realtime video editing (add/replace/remove objects, swap characters, change backgrounds live) $0.02/s at 720p or $0.04/s in fast mode, $0.04/s for batch video; Lucy Restyle 2 $0.01/s; image ...
• articleOpen interactive world models you can run or train: Matrix-Game 3.5 (720p, long-horizon memory, first/third person, 40 GB GPU) and ForgeWM (fully open recipe, real-time keyboard/mouse control, trained on 8 GPUs)
  Two open releases extend the open world-model landscape (C3-B399-01). Matrix-Game 3.5 (Riemann Dynamics; HF weights apache-2.0, July 2026; repo Sep 2026) (R1, R2): memory-augmented interactive world model for 720p long-horizon camera-controllable video in first-person and third-person modes. "Patch Memory" lifts past observations into a 3D memory with geometric alignment and visibility-aware retrieval so revisited places stay consistent; static scene and dynamic subjects are separated, with third-person mode taking 0-4 protagonist reference crops for identity. Built on Wan2.2-TI2V-5B with Depth-Anything-3; 5B base (bidirectional) checkpoints plus a three-step distilled first-person causal checkpoint for real-time streaming. Needs one NVIDIA GPU with >= 40 GB VRAM (704x1280 peaks around 40 GB), Linux, >= 64 GB RAM. ForgeWM (arXiv Aug 2026) (R3, R4): an end-to-end open recipe (bidirectional SFT -> teacher-forced causal AR -> consistency distillation -> on-policy DMD) producing 1/2/4-s...
• articlebvhio (addb5461, MIT) accepts exactly one ROOT hierarchy followed by MOTION; the informal Wisconsin description notes BVH lacks a full basis-pose definition. No formal BVH specification was found by the...
  bvhio (addb5461, MIT) accepts exactly one ROOT hierarchy followed by MOTION; the informal Wisconsin description notes BVH lacks a full basis-pose definition. No formal BVH specification was found by the producer.
• articleglTF 2.0 perspective cameras carry only aspectRatio, yfov, zfar and znear, so converting a physical camera (focal length, apertures, offsets) to glTF keeps only field of view and ratio; recovering focal length from glTF needs sensor size supplied out of...
  glTF 2.0 perspective cameras carry only aspectRatio, yfov, zfar and znear, so converting a physical camera (focal length, apertures, offsets) to glTF keeps only field of view and ratio; recovering focal length from glTF needs sensor size supplied out of band.
• articleretarget_bvh (8501d53f) writes rotation keys for mapped bones and location keys only for the parentless (hip) bone, so bone translations other than the root are not retargeted; the add-on manifest declares GPL-2.0-or-later and Blender 4.2.0 minimum (the wiki says 3.0+, a version-statement...
  retarget_bvh (8501d53f) writes rotation keys for mapped bones and location keys only for the parentless (hip) bone, so bone translations other than the root are not retargeted; the add-on manifest declares GPL-2.0-or-later and Blender 4.2.0 minimum (the wiki says 3.0+, a version-statement conflict per producer).
• articleOpenUSD v26.08 usdrecord defines --abortOnError (default False); without it, rendering Tf errors do not by themselves force a non-zero exit, while specific failures (e.g. stage open) return 1 (static reading; producer reports default exit 0 on Tf...
  OpenUSD v26.08 usdrecord defines --abortOnError (default False); without it, rendering Tf errors do not by themselves force a non-zero exit, while specific failures (e.g. stage open) return 1 (static reading; producer reports default exit 0 on Tf errors).
• articleAssimp's FBX importer (c3f68d29) declares supported FBX versions 7100-7400 ('FBX 2011, FBX 2012 and FBX 2013'); newer files fall outside that window and are handled only as the code's newer-version path allows (producer: warning unless...
  Assimp's FBX importer (c3f68d29) declares supported FBX versions 7100-7400 ('FBX 2011, FBX 2012 and FBX 2013'); newer files fall outside that window and are handled only as the code's newer-version path allows (producer: warning unless strictMode).
• articleWhat do RenderMan 26's public docs say about Points primitives (shape, falloff, width, camera-facing) that matters for rendering Gaussian splats as points?
  Corrected 2026-10-23 - see C3-B643-03; the original statements below are kept for history and are superseded where the correction says so. RenderMan 26's public docs describe points as disks (with normals) or real spheres (without), sized by constantwidth or per-point width (default 1.0, object space), and list a falloffpower primvar (default 0) for point primitives. None of the pages read documents points:shape, falloffshape, camerafacing, Gaussian falloff or splat rendering. So the override values hdPrman uses for Gaussian splats (seen in earlier reviews) can't be explained from public RenderMan docs.
• articleWhen exporting a simulated garment from Marvelous Designer, what do the Alembic and cache-animation (MDD, PC2, Maya cache) exports include, and what do the help pages leave out?
  Marvelous Designer hands cloth on as per-vertex motion: Alembic (HDF5 or Ogawa) or garment-only caches (Maya .mc/.mcx, Point Cache 2, MDD in two variants). Export options cover weld vs unweld seams (which changes vertex counts and object splitting), thickness, merged vertices, unified 0-1 UVs, optional avatar shape, scale, axis conversion and frame range; strain or stress as vertex colour is documented only for Alembic. Pattern, physical-property and sewing data travel only as an optional XML with the cache export. The pages do not say how fabric materials or simulation settings reach a renderer.
• articleOverview: for capture to render, what does each file or API in the chain actually carry (markers, blendshape weights, skeleton animation, caches, render products), and what is not carried?
  Synthesis (producer, supported by verified statements): each capture-to-render carrier defines its own fields but not the mapping to the next: C3D labelled markers, ARKit/MediaPipe 52 named coefficients, UsdSkel joints with world-space bind transforms, Alembic time-sampled geometry whose topology may change between samples, UsdRender settings with no universal RenderVar names, and EXR named channels whose layer naming implies no stacking order. None carries a rig, a character or an intent, so marker-to-joint, blendshape-name and RenderVar-to-channel mappings must be authored and recorded.
• articleField by field, how do the ratified KHR_gaussian_splatting extension and OpenUSD's ParticleField3DGaussianSplat schema define a splat's position, rotation, scale, opacity, spherical harmonics, kernel, projection, sorting, colour space and frame, and where do they differ?
  Both definitions store linear scale and linear 0..1 opacity (CA-B035), but defaults and fields differ: glTF KHR_gaussian_splatting requires a colorSpace string and defaults sortingMethod to cameraDistance, while the OpenUSD usdVol schema defaults sortingModeHint to zDepth and has no colour-space attribute, so a converter that copies defaults changes the declared sort metric and loses colour-space provenance. The glTF README carries an SPDX CC-BY-4.0 header and also a Khronos copyright statement limiting reproduction to the unmodified specification.
• articleWhat does the Depthflow ComfyUI node pack provide for 2.5D parallax, what does it pin, and what licence does it carry?
  Problem. Making camera-move parallax animations from a still plus a depth map without a video model. Conditions. akatz-ai/ComfyUI-Depthflow-Nodes commit 1806b6ed (pushed 2026-02-15; not archived), version 1.2.3, licence AGPL-3.0 (LICENSE file and API). Method. The Depthflow node takes an image (or batch) plus a corresponding depth map and applies motion objects (zoom, dolly, circle and others) with optional stackable effects (depth of field, vignette), outputting an image batch to be encoded as video; parameters include num_frames, input_fps/output_fps, quality, ssaa and invert, with tiling_mode. It extends RyanOnTheInside's Flex system. pyproject and requirements pin depthflow==0.9.1 plus pydantic, numpy and opencv-python. The README credits the upstream DepthFlow library by Tremeschin (BrokenSource) and publishes a custom Docker environment image; node documentation lives on an external Notion page. Failure modes. Depth-map quality drives artefacts (the depth source is outside thi...
• articleWorld Labs Marble (Oct 2026): persistent 3D worlds from text, images, video or a 3D blockout - models, credit costs, plan rights and export paths
  Primary docs. Marble generates persistent 3D worlds (Gaussian splats, optional meshes) from a text prompt, single image, multiple images, 360 panorama, video, or a coarse 3D layout built in "Chisel" (draw walls, doorways, place a panorama camera, or upload GLB/FBX geometry, then prompt). Studio tools: Compose (join several worlds into larger environments), Record (camera flythroughs exported as video), pano edit, Expand (fix low-fidelity regions/grow worlds), variations. Models (R1): Marble 1.1 (recommended, 1,500 credits per world), Marble 1.1 Plus (largest worlds; 1,500 base + 0-1,500 variable), Marble 1.0 legacy (1,500), Marble 1.0 Draft (150 credits, fast tests) - 1.1/1.1 Plus added 2026-04-02 (R2). Credit examples (R3): single-image world = 1,500 + 80 input credits; edits 150 per pano edit; expand 2,000. Plans: Free (text/single image/pano, up to 4 worlds), Standard (multi-image, video and 3D-layout inputs, editing, export; ~12 worlds), Pro (expand, enhanced video output, high-...
• articleWhat planning framework and location-consistency method does a solo AI animation workflow use (SCENE, 3D world from images, shot-by-shot prompts)?
  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. Youri van Hofwegen (2026-05-05, OpenArt affiliate) plans before generating with "SCENE": Story (inciting incident; decide the ending first), Characters (who and rough look), Emotion (decide the palette/mood arc up front, e.g. tension -> danger -> relief), Narrative beats (three clips = setup, escalation, resolution; one job per clip because the model does not know the next clip), Every-clip rules (a fixed style line in every prompt as a "production bible"). Location consistency: OpenArt's "3D World" builds a navigable world from 1-4 multi-angle images of a location (generate angle variants first, "Ultra" mode), then "take shot" anywhere with focal length 23-300 mm; char...
• articleHow was a large-scale ocean/kaiju water shot simulated in Houdini (particle counts, wedging), as a traditional VFX reference point?
  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. In a short making-of (2026-01-16) Mikkel Lassalle reports his kaiju ocean shot used a Houdini FLIP simulation of about 50 million particles plus about 70 million whitewater particles; to make the whitewater feasible he split the FLIP tank into four wedges, simulated each separately, and combined them at render; the environment was matte paintings composited in After Effects with "simple blend modes and color corrections" (method transferable to Nuke/Fusion); HUD graphics also in After Effects; creature by a collaborating 3D artist, music/sound by another collaborator. Value: a concrete scale/time reference to contrast with AI-generated water "for free" in layout-to-rend...

More