Evaluation and media provenance

Benchmarks, task datasets, attribution, metadata, interoperability and quality review

• articleWhat do the vendors' terms and help pages say about using user inputs and outputs to train models, by plan or surface, and about restrictions on using outputs to build competing models?
  Training statements vary by surface. OpenAI: consumer content may be used to train with an opt-out; API and business content is not used by default. Google Cloud: will not use Customer Data to train without permission. Adobe: says it does not train on Creative Cloud subscribers' personal content and that partner models do not change that. Midjourney and Ideogram contain no clear 'we train on your content' or 'we do not' sentence in the terms read; Midjourney's privacy policy lists data collected through training its algorithms and grants itself a perpetual licence over inputs and assets; Ideogram's privacy policy lists training the models among uses of information collected. Google Cloud and OpenAI bar using outputs to build similar or competing models, and Midjourney bars using the Services to develop competitive products.
• articleAt the observation dates, do OpenAI, Google Cloud (Imagen), Midjourney, Ideogram and Adobe Firefly say the user owns image outputs and may use them commercially, and does the plan matter?
  OpenAI (consumer terms and business agreement), Ideogram and Google Cloud state in their terms that the vendor does not claim ownership of outputs (OpenAI and Ideogram also assign any rights they hold; Google says Generated Output is Customer Data and it asserts no ownership). Midjourney says the user owns assets but ties ownership to a Pro or Mega plan for companies above one million USD annual revenue, and its help article words the same rule as needing the plan to use images commercially. Ideogram's help FAQ defers commercial use to its terms, while the terms do not restrict it; its plan matters for privacy of images. Adobe's FAQ permits commercial use of non-beta Firefly output but is silent on ownership in the text read. This candidate records texts, not legal conclusions.
• articleWhat does WorldMemBench's fixed-reference multi-revisit protocol measure, what did it find about decay across returns, axis switches and off-screen events, and how reliable are its checks?
  WorldMemBench (260 cases: 200 static with human, object and environment targets, 60 dynamic with off-screen events) compares every prescribed return with the same first-visit reference over translation loops and yaw/pitch cycles, giving a score per revisit and counting a missing target as zero. On nine open video world models, the static leader (HY-WorldPlay) initiates only 31.7% of prescribed events, scores decay from the first to the third return on all eight metrics, and switching from a yaw cycle to a pitch cycle costs more. The paper adds pose-qualified subsets to separate memory from bad camera execution, and a context-policy study on LingBot World v2. Human agreement is 78.5-92.5% for static metrics but model-level correlation varies from 0.301 to 0.988.
• articleWhich of the explicit-memory world-model papers and revisit benchmarks can I actually download or run, and what do their licence texts, metadata and gates say per component?
  Observed on 2026-10-22: WorldMem has code (non-commercial licence text) and public ungated checkpoints with no licence field; SPMem has Apache-2.0 code and Apache-labelled checkpoints; VMem has MIT-labelled code that vendors a CC BY-NC-SA component and weights behind a Hugging Face access gate; Context-as-Memory has a released dataset but no code I could find; WorldTrace has no code I could find. The two newest benchmarks (R2M-Bench, WorldMemBench) release MIT code and data, but WorldMemBench's dataset card says its third-party images and the generated videos keep their original rights, and its examples set is labelled CC BY-NC by metadata. Nothing here was run, so 'released' means files are listed, not that they work.
• articleWhat must a custom character satisfy to be used in Flow Studio, which terms govern its outputs, and how have the documented capabilities changed over recent dated releases?
  The character requirements page (stamped Oct 19, 2026) sets file, polygon, hair-strand and DCC-version limits, T-pose and add-on validation requirements. The release notes show the platform broadening in 2026 from live-action replacement into a 3D Editor and a Canvas that hosts third-party generative video/audio models, so "Flow Studio" now covers more than the original tool. Output ownership is not stated on the pages read: the only terms linked from the help site are Autodesk's General Terms (Internet Archive snapshot 2026-10-16), which say users keep ownership of what they submit, define "Output", and disclaim responsibility for it; product-specific Special Terms were not read, so Flow Studio output rights remain unresolved.
• articleWhere do the licence layers of an open V2A model disagree or stack (ThinkSound's Apache vs non-commercial text, Stability's VAE licence, VGGSound/AudioSet/BBC training data)?
  ThinkSound publishes four incompatible-looking statements: HF weights card Apache-2.0 (S15, S16), GitHub README Apache-2.0 headline but research-only/no commercial use (S17), a bundled Stability AI Community License for the fine-tuned VAE (S18; commercial use needs registration and a revenue condition), and a paper statement that the dataset is strictly academic/non-commercial (S14). Its training sources include VGGSound (page: CC BY 4.0, copyright remains with video owners), AudioSet (CC BY 4.0 annotations for YouTube-linked data), Freesound and BBC Sound Effects (site: personal/educational use or license). A permissive label on one file does not resolve the others.
• articleWhen a model's GitHub LICENSE, Hugging Face licence metadata and model-card prose disagree, does Hugging Face or the vendor say which governs, and what does the HF licence field officially mean?
  At the pinned revisions I found no Hugging Face page and no vendor text (Whisper, VibeVoice, AudioSeal) that ranks the GitHub LICENSE, the HF front-matter licence and the card prose against each other. HF's docs describe the licence field as the owner's way to 'let other users know' the permissions, shown on the model page and used for filtering; custom terms belong in a LICENSE file in the repo. HF's Terms of Service say content carrying a licence notice is intended to stay under that licence and that downloaded content is subject to the Terms 'and/or' the terms accompanying it. So the safe agent reading is: record each channel separately with its revision, call the HF field 'metadata says', and do not pick a winner.
• articleDo the Cosmos repositories, model cards and docs restate the NVIDIA Open Model License accurately, and do the dated versions that appear in model-card front matter differ from the current text?
  Three dated agreement texts are readable without accepting anything: the current page (October 24, 2025), the text embedded in the Cosmos-Predict2.5-14B card (September 23, 2025, an abridged rewording), and the text embedded in the Cosmos-Tokenize1 card (January 6, 2025). The January text uses 'Model Derivatives', has no Special-Purpose Model clause, no 'Guardrail' definition and no Separate Components section, and its guardrail sentence has no 'substantially similar Guardrail' exception. GitHub READMEs and the docs licence page restate the licence only by name and link. One Cosmos 3 model repository ships a LICENSE file that matches the October 2025 text. Model-card 'Important Note' paraphrases of the guardrail clause differ between cards.
• articleBefore I use, host, fine-tune or redistribute a Cosmos model that is under the NVIDIA Open Model License, what does each clause of the current agreement text say, and which of my questions does it not answer?
  At the dated text 'Last Modified / Version Release Date: October 24, 2025' the agreement has 11 numbered sections. It grants a perpetual, worldwide, no-charge, revocable licence to use, modify, make derivatives of, sell and distribute the Model; allows redistribution with a copy of the agreement and a Notice line; requires 'Built on NVIDIA Cosmos' for Cosmos Models in several uses (including training another AI model with them or their outputs); states NVIDIA claims no ownership of outputs; and ends your rights automatically on litigation or on bypassing a Guardrail without a substantially similar one. It has no clause on hosting as a service by that name, no express patent licence (the word 'patent' appears only in the litigation-termination sentence) and no operative commercial-use clause beyond the preamble statement. This is a reading of text, not legal advice.
• articleThe agreement ends my rights if I bypass a Guardrail without a substantially similar one. What do the agreement text, the Cosmos docs and the Cosmos code at a pin say about whether the guardrails can be turned off?
  Three source kinds disagree on tone, not necessarily on fact: the agreement (October 2025) ends rights if you bypass a Guardrail without a substantially similar one; the Cosmos docs state the Cosmos-1.0-Guardrail model 'cannot be disabled' in the diffusion and autoregressive pipelines; and cosmos-predict1 at one commit defines a --disable_guardrail command-line argument and forwards it to the pipeline. The docs sentence and the code are about overlapping but not proven-identical pipelines; the code path was not traced or run. 'Substantially similar Guardrail appropriate for your use case' is not defined in the text.
• articleDid the AudioSeal vendor ever publish a separate weights licence file, what do the GitHub and Hugging Face histories show, and how can an agent check release-scoped weights terms?
  Yes, in an earlier revision. At git 785b3b11 the AudioSeal repo has a LICENSE_weights file carrying Creative Commons Attribution-NonCommercial 4.0 text and a README that says code is MIT and weights CC-BY-NC 4.0. Commit 29565154 ('Re-update of license (#28)', committer date 2024-04-02) removes LICENSE_weights and edits the README; at HEAD e63a8a0e the only LICENSE is MIT and the weights are covered only by a README news line, while the README License section still names only code. The Hugging Face repo has no licence file in its tree, and its README gained license: mit in a later commit than its first two README versions.
• articleHow does Physics-IQ test physical understanding in video models, which models and versions were scored, what were the results, and what do the authors say they do and do not show?
  Physics-IQ (Motamed et al., Google DeepMind) gives a model the first 3 seconds (or the last frame) of a real, static-camera physics event and asks it to continue for 5 seconds, then compares to the real continuation with four motion-based metrics, combined and normalised so two real takes of the same scene score 100. At v3 the best of eight configurations (VideoPoet, multiframe) scores 29.5 and Sora (image-to-video) 10.0; visual realism judged by Gemini 1.5 Pro was not significantly correlated with physics score (r = -0.46, p = .249). The authors call the metrics proxies, say they are on the conservative side (they penalise camera moves and shot cuts), and say results do not show that future models cannot learn physics.
• articleWhere does a Hugging Face model repository tagged creativeml-openrail-m keep its licence text, and how do I check what it contains?
  The CompVis stable-diffusion-v1-4 card (133a221b) declares creativeml-openrail-m, allows redistribution and commercial use only with the same use restrictions passed on, and points to a Hugging Face Space for the full text, with no LICENSE-named file at the repository root; under the same metadata value one derived repo ships the CreativeML Open RAIL-M text (dated 22 Aug 2022) as LICENSE.md, while pt-sk/stable-diffusion-1.5 ships an unfilled MIT licence template as license.txt.
• articleUnder what terms may an agent use AudioSeal code and checkpoints, and how does the repo's open detector sit with its own paper's advice?
  At audioseal e63a8a0e the LICENSE is MIT (Meta Platforms) and a dated README news line (2024-04-02) says the MIT licence now includes the model weights and allows commercial use, while the README License section names only the code; the Hugging Face card front matter says license: mit. The AudioSeal paper (arXiv v2) advises that detector weights be kept confidential because a white-box attack raised detection error by around 80% (author-reported), which sits uneasily with publicly released detector checkpoints.
• articleWhat do the Llama 3.3 and Llama 4 Acceptable Use Policies prohibit, and how do the community licences they are incorporated into treat redistribution, modification, training on outputs and hosting?
  At llama-models 0e0b8c51, the Llama 3.3 and Llama 4 terms permit redistribution and derivative models on conditions (copy of the agreement, 'Built with Llama' display, Notice file, 'Llama' at the start of the name of any distributed model trained or improved with the materials or their outputs; separate licence above 700 million MAU), and both AUPs prohibit representing the model's use or outputs as human-generated. For multimodal models, the Section 1(a) grant is withheld from EU-domiciled individuals and EU-based companies, but this does not apply to end users of products incorporating those models.
• articleWhat do the Gemma Prohibited Use Policy and Gemma Terms of Use (April 2026) say about hosting, redistribution, modification, training on outputs and output use, and how does the Gemma 4 licence differ?
  The Gemma Terms of Use (last modified 1 April 2026) treat offering Gemma or a Model Derivative as a hosted service as Distribution, count models trained on Gemma's synthetic outputs as Model Derivatives (outputs themselves are not), allow reproduction and distribution only with the use restrictions passed on as enforceable terms and other conditions, and state Google claims no rights in outputs; the incorporated Prohibited Use Policy (21 Feb 2024) bars passing generated content off as human-made to deceive. The terms page points Gemma 4 to separate terms, and the linked page is an Apache License 2.0 page.
• articleWhen a model repository ships code, weights, data references and generated output, what do the licence texts themselves say about each component, and where do the records disagree?
  Under the OpenRAIL-M text, code/model/derivatives, outputs and training data are addressed by different clauses, and training Data is carved out of the licence; a single "licence: openrail" field cannot stand for all components.
• articleCan I reuse govuk-frontend, USWDS or Carbon source, and what do the licence files at the pinned tags actually say?
  At the pins read, govuk-frontend is MIT (Crown Copyright GDS 2017) and Carbon's root LICENSE is Apache-2.0, while USWDS's LICENSE.md is a mixed notice: GSA modifications are public domain/waived, but bundled fonts (SIL OFL 1.1), Material Icons (Apache 2.0) and the sprite carry their own terms, so reuse rights depend on which paths are copied.
• articleWhat does the VRM Public License 1.0 text itself state, and how does it hand control to the per-model "License Settings"?
  The VRM Public License 1.0 (English page at vrm.dev/en/licenses/1.0/) grants Model Use and makes other rights depend on the License Settings stored in each VRM file; the alternative URL form vrm.dev/licenses/1.0/en/ returned 404 to curator GET, consistent with the producer observation.
• articleWhich machine-readable licence and usage fields does VRM 1.0 (VRMC_vrm.meta) define, and how do they differ from VRM 0.x meta?
  VRM 1.0 meta requires name, authors and licenseUrl and defaults to restrictive values (onlyAuthor, personalNonProfit, credit required, modification prohibited, redistribution false), so an absent field is not a permission; VRM 0.x field names keep the misspelling "Ussage", which parsers must match literally.
• articleDo measured studies support 'bigger text and looser line spacing are always more readable', and does a specialised dyslexia font help?
  Katzir et al. (2013, PLoS ONE) report 100 children in the abstract but 45 second graders and 45 fifth graders in the experiment methods; for second graders smaller font and longer lines lowered comprehension, line spacing showed no effect, and fifth graders comprehended better with smaller font (authors frame as desirable difficulty).
• articleWhat do FADGI video resources specify or compare for video reformatting targets (RDD 48 / AS-07, wrapper and encoding matrices)?
  FADGI describes RDD 48 (from AS-07) as a vendor-neutral MXF archive/preservation subset with a Baseband Shim annex and lists the current version as dated 3 December 2018; IASA-TC 06 Part B (2019) says RDD 48 was published in 2019 and is largely derived from AS-07.
• articlepython-soxr (351f265b) defaults to quality 'HQ' in both its one-shot resample() and its streaming class (curator read), so 'soxr default' in Python differs from libsoxr's C one-shot default reported as LQ by the...
  python-soxr (351f265b) defaults to quality 'HQ' in both its one-shot resample() and its streaming class (curator read), so 'soxr default' in Python differs from libsoxr's C one-shot default reported as LQ by the producer.
• articleSPDX 3.0.1 AI profile conformance requires exactly one hasConcludedLicense and exactly one hasDeclaredLicense relationship per AIPackage; licences attach via relationships, not a package property; the spec defines declared as what the artifact was discovered to contain (e.g. by tooling) and...
  SPDX 3.0.1 AI profile conformance requires exactly one hasConcludedLicense and exactly one hasDeclaredLicense relationship per AIPackage; licences attach via relationships, not a package property; the spec defines declared as what the artifact was discovered to contain (e.g. by tooling) and concluded as the SPDX data creator's conclusion.
• articleAt diffusers 086bf957 the guide recommends a CPU Generator for reproducibility, and the shipped enable_full_determinism() hard-codes CUBLAS_WORKSPACE_CONFIG=':16:8' while the repository's own test helper switches to ':4096:8' for torch >= 2.10 citing...
  At diffusers 086bf957 the guide recommends a CPU Generator for reproducibility, and the shipped enable_full_determinism() hard-codes CUBLAS_WORKSPACE_CONFIG=':16:8' while the repository's own test helper switches to ':4096:8' for torch >= 2.10 citing CUBLAS_STATUS_NOT_INITIALIZED.
• articleAt hub-docs 801da850 the licence table lists 'unknown' and 'other' and routes license: other to license_name plus a LICENSE file; the live licence-tags endpoint returned 82 ids without 'unknown' (curator GET 2026-10-21); huggyllama/llama-7b declares only 'other' with a LICENSE file present, so...
  At hub-docs 801da850 the licence table lists 'unknown' and 'other' and routes license: other to license_name plus a LICENSE file; the live licence-tags endpoint returned 82 ids without 'unknown' (curator GET 2026-10-21); huggyllama/llama-7b declares only 'other' with a LICENSE file present, so restrictive terms are reachable only by reading the file.
• articleWhat check can an agent run on a draft or outline against an intended audience before production, built from Baker (1919, Chapter X) and labelled 1919 stage practice and a proposal?
  PROPOSAL built on 1919 stage practice. Baker gives no checklist; this batch assembles one from Chapter X (labelled synthesis), in this order: (0) freeze the central human and artistic purposes; (1) state the intended audience, its knowledge and approach, and whether it is general or special; (2) interest check; (3) comic/tragic and sympathy check; (4) motivation check for informed viewers; (5) form and length check; (6) order check; (7) truckling triage: any audience-driven change that alters the purposes is refused, and an audience-unfriendly element is cut only after a recorded attempt to make the audience feel it; (8) record each change with reason and stage; (9) readiness gate. It is run as a separate critical pass, in line with Baker's advice to keep creative and critical work apart. Baker supplies no outside-audience protocol; any LLM or human reader panel is an untested addition.
• articleHas the OSSF model-signing-spec conformance suite changed after commit 91bf841e8ccf97a3ed97309387e9298d1ac18ba2, and are the curator's two roundtrip.yaml / models.yaml inconsistencies fixed?
  Observed 2026-10-22T14:19:26Z, the GitHub commits API for ossf/model-signing-spec filtered to path=conformance lists exactly three commits on the default branch, and the newest is 91bf841e8ccf97a3ed97309387e9298d1ac18ba2 (2026-09-18T14:53:02Z), the commit the curator used. So no later commit changed conformance/ and both files are unchanged from the pinned versions: the stale .github comment still conflicts with spec 6.2 (which lists .github), and subdir/.gitignore is still typed signed in with-nested-gitignore and ignored in multi-gitignore. I did not run the suite and did not re-fetch the files; the excerpts reuse the cycle-2 pinned copies whose hashes match.
• articleWhat exactly does Adobe's Illustrator generative-AI FAQ say about when Illustrator adds, preserves or does not add Content Credentials, by Illustrator version?
  At the fetch time (2026-10-22T13:51:25Z) the page, whose own date line reads 'Last updated on Jun 1, 2026', states one rule for Illustrator 29.3 and earlier (Content Credentials added to all new generative AI content, and imported ones preserved) and a different, conditional rule for 29.4 and later (no addition when the file has no generative AI content; imported Content Credentials preserved on export). The 29.4-and-later sentence is quoted below as printed; the page gives no definition or example of 'contain generative AI content' in that item. This candidate reports what Adobe states, not how the product behaves.
• articleWhat do ASCAP, BMI, SOCAN and PRS for Music publicly say about AI-generated music in registration, and do library pages say anything about it?
  A joint release dated 2025-10-28 (hosted on SOCAN's site) says ASCAP, BMI and SOCAN each adopted policies to accept registrations of partially AI-generated compositions, defined as combining AI-generated musical content with human authorship, and that entirely AI-created compositions are not eligible. BMI (403), ASCAP's own AI help page (403) and a PRS AI policy PDF (404 at the URL a search returned) could not be read, so the societies' own policy texts and PRS's position are not inspected here. None of the library, publisher or Guild pages read contains any AI statement. The release says nothing about cue sheets, clearance or licensing of AI-made music.
• articleWhat is documented about platform AI policy toward authors in Chinese web fiction, and what does platform economics say about serial incentives?
  Thin, secondary. One news item (AsiaIPLaw, 26 July 2024) reports that ByteDance's Fanqie Novel added an AI-training supplement to author contracts (works usable for training, platform copyright over generated content, priority beta access) and offered a termination feature on 23 July 2024; it also states the platform's revenue share rewards number of readers and reading time. The Rest of World 2026 article on platform restrictions of AI-generated fiction could not be fetched (HTTP 429), and Qidian/Webnovel AI rules were not found; later 2026 developments appear only in search summaries and are not evidence here. No source read gives chapter length, update cadence or hook rules.
• articleWhat does The Chosen One (iterative consistent-character extraction from a text prompt alone) cost per character, how was it measured, what failure modes do the authors list, and is there official code?
  The Chosen One (arXiv 2311.10093v4, SIGGRAPH 2024) takes only a text prompt, generates a set of images, clusters their embeddings, optimises LoRA weights and two learned text tokens on the most cohesive cluster, and repeats until convergence. The authors report 1-2 iterations, about 13-26 minutes on an A100 (bfloat16), and list the cost as a limitation (about 20 minutes). Consistency is measured as pairwise CLIP image-embedding similarity across prompts and prompt similarity as CLIP score; the main result is a trade-off plot (Fig. 6) plus an AMT user study; the numbers are not in the HTML text. Stated failure modes: identity may not fully converge, a supporting character such as a pet is not kept consistent, spurious attributes can bind to the identity, one character at a time, and simplistic centred scenes. The author project page lists only an 'unofficial' code link. The catalogue already holds this paper as a source entry; this candidate adds the pinned v4 details.
• articleWhich classic craft concepts (key, breakdown, in-between, spacing, holds) do the AI in-betweening tasks actually correspond to, and is any production use documented?
  Editor's mapping (synthesis, not stated by any source): the papers' "given frames" are Lutz's extremes/keys; the generated frames are in-betweens; ToonComposer's optional middle sketch plays the role of a middle (breakdown-like) key; spacing is controlled only implicitly (sketch positions, FPS conditioning, text) and holds and exposure (on twos) are not modelled in either ToonCrafter or ToonComposer (no hit for hold, twos, spacing, timing, breakdown or exposure in either paper body). No production use is documented in the sources read: ToonComposer thanks two studios for permission to use footage in evaluation; ToonCrafter calls itself a research exploration. Production use is therefore a hypothesis here.
• articleWhat do the earlier AI in-betweening papers (AnimeInterp 2021, AnimeInbet 2023) and AniDoc (2024) claim about inputs, failure modes and data, and what licences do their code, weights and datasets carry?
  Abstract- and README-level only (full papers not read for storage reasons). AnimeInterp: raster-frame interpolation of one in-between from two frames, framed around textureless flat regions and large non-linear motion, with the ATD-12K triplet dataset. AnimeInbet: vector/line-graph inbetweening of black-and-white line drawings, because raster warping blurs lines; ML240 dataset of 244 sequences; Chamfer distance code. AniDoc: video-diffusion colourisation of sketch sequences, with in-betweening from start and end sketches. Terms: AnimeInterp code MIT by README only; AnimeInbet dataset CC BY-NC-SA 4.0 and code non-commercial; weights and AnimeInterp data links carry no terms in the READMEs.
• articleWhat does ToonComposer (2025) claim for AI "post-keyframing" from sparse keyframe sketches, what was measured on real artist sketches, and what are its failure modes and licences?
  ToonComposer (arXiv v1, Aug 2025) merges in-betweening and colouring into one step on a Wan2.1 video model: a colour reference frame plus one or more keyframe sketches (any frame position) and optional region masks produce the clip. Evaluation: synthetic benchmark from cartoon films used with permission, PKBench (30 scenes, start/end sketches drawn by professional artists) and a 47-person preference study (70.99% aesthetic, 68.58% motion for ToonComposer). Stated problems: baselines distort faces and are sensitive to bold line styles; blank sketch regions render flat unless region masks are used; compute cost (README: 61 frames at 480p needs ~57 GB VRAM). Licence statement is MIT including weights.
• articleWhat does ToonCrafter claim and measure for generative cartoon in-betweening (inputs, frame counts, metrics, user study, failure modes) and what are the code/weights/data terms?
  ToonCrafter (arXiv v1, May 2024; SIGGRAPH Asia 2024 journal track per README) generates intermediate frames between two cartoon images with a fine-tuned image-to-video diffusion model, optionally guided by sparse sketches. Released model: 16 frames at 512x320 from a start and end image plus text. On its own 1K cartoon test set it reports FVD 43.92 vs 98.96 for the next best (SEINE) and a 24-person study with 68.98% preference for motion quality; it does not win LPIPS (0.1733 vs 0.1702 FILM). Stated failure modes: baselines fail on dis-occlusion and non-linear motion; unadapted priors leak live-action content; compressed latents cause flicker/detail loss; the maintainers state the success rate is not guaranteed.
• articleWorldMark reports a 'Response Latency' for ten released interactive world models: what does it measure, is it compute or wall-clock latency, what are its values, and how do they relate to throughput figures?
  WorldMark v2 (Alaya Lab and others; not the model authors) scores 'Response Latency' as how soon the generated video's optical-flow speed reaches half its normal sustained speed after a command switch, normalised to 0-100 (higher is better). It is a content-response metric, not compute time: the paper reports no generation time or FPS, only that scoring one model on one split takes about one hour on two H200 GPUs excluding video generation. First-Person Real translational/rotational Response Latency: Matrix-Game 2.0 98.23/86.47, LingBot-World 86.56/99.20, HY-World 1.5 85.44/85.13, HY-GameCraft 1.0 90.97/51.99, Yume 1.5 40.66/77.11. Yume 1.5 has the lowest translational Response Latency here yet the highest Hz in WorldRoamBench, so the two measure different things. The v1 HTML has six models and no latency metric.
• articleWhat non-author evidence exists on the time from starting an interactive world model to the first usable frame, which parts of that cold-start time are separable (loading, compile or kernel warm-up, offload transfers), and which were fixed upstream?
  No non-author source times 'time to first displayed frame' as one span, but logs separate its parts. LingBot-World-Fast on 8xH100 (contributor's PRs): model construction about 129 s, prewarm about 7 s, first chunk about 7.2 s versus about 1.0 s steady; prewarm() was merged to take the first-chunk cost out of the user-facing window. HY-WorldPlay Wan-5B on 4xA100 (user): first four chunks 89-368 s, then about 3.4 s; a maintainer attributes it to torch.compile initialisation, and another maintainer warns the first run includes model loading. Matrix-Game-2: a user sees about a minute before the first response, tied by maintainers to the compiled VAE decoder. Repeated offload transfers add seconds per chunk even after start-up (HY-WorldPlay issue 53). Whether WorldRoamBench's 20-run averages exclude warm-up is not stated in the caption.
• articleWhich released interactive world models have a speed or latency figure measured by someone other than their authors, where is it, and which systems have none?
  Non-author timings exist for several systems but live in only four kinds of place: a benchmark's efficiency table (WorldRoamBench v5 times 10 open models on one H20), an efficiency paper's baseline row (QuantWM v3 times LingBot-World-v2 and HY-World 1.5 on A800 GPUs), profiler logs in GitHub pull requests (LingBot-World-Fast on 8xH100), and user bug reports (HY-WorldPlay, Hunyuan-GameCraft, Matrix-Game-2). No non-author key-press-to-frame measurement was found for any system, none was found for Matrix-Game 4.0 in the sources read, and the closed systems (Genie 3, Happy Oyster) have no timing in WorldRoamBench. WorldMark's 'latency' is an in-video response metric, not wall-clock time.
• articleWhat do TikTok and Meta say creators must label as AI at upload, which labels do the platforms add themselves, and which parts are announced but unverified?
  TikTok (Support page via the IA capture of 2026-06-27; Newsroom post of 2025-11-19): creators are encouraged to label completely generated or significantly edited content and are required to label realistic AI content; routes include captions, hashtag stickers or the "AI-generated content" post setting. Misleadingly labelling unaltered content is a Terms violation. TikTok may also apply an automatic label (TikTok AI effects, uploaded Content Credentials) which the creator cannot remove, and says detection models back the rule. Invisible watermarks and a viewer AI-content control were announced in November 2025 and not observed. Meta: 2024 posts say Meta would require its disclosure tool for photorealistic video and realistic audio (penalties possible), and adds "AI info" labels when it detects industry-standard indicators or when people self-disclose; from September 2024 labels for content only edited by AI moved to the post menu. Meta's current Transparency Center page rendered no ...
• articleWhat does Steam's Content Survey ask about AI-generated content, what changed in its wording, and what stays inside Steamworks behind login?
  Steamworks documentation (public page, captured 2026-10-22) says the Content Survey has a third section on generative AI that the developer must complete before review. It separates Pre-Generated content (shipped content created with AI tools during development) from Live-Generated content (created while the game runs), and for live-generated content the developer must describe guardrails against illegal output. Valve says it compares survey answers with the build and store page, and that the survey does not replace the content rules. Internet Archive captures show the Pre-Generated definition read "Any kind of content (art/code/sound/etc) created with the help of AI tools during development" at least until 2026-05-13, and by 2026-06-17 it was scoped to content "that ships with your game and is consumed by players", with an added sentence that efficiency gains from built-in AI tools are not the focus. A secondary report dated a rewrite to 16 January 2026; the official page captured ...
• articleHow do DistroKid, TuneCore and CD Baby treat AI-generated or AI-assisted music at upload: do they accept it, ask for a disclosure, or exclude it?
  Three different models appear in the Help Center text read on 2026-10-22. DistroKid accepts AI-tool music on rights and impersonation conditions and asks at track upload whether any of it was AI-generated (lyrics, music, all of the audio, part of the audio; an "All of the audio" answer triggers a human-or-AI-persona question; credits can be edited later; pure tool use such as pitch correction or AI-assisted mastering needs no credit). TuneCore distributes AI-assisted music only if every GenAI tool used was trained on fully licensed datasets, and its publishing arm says it runs AI detection and that songs flagged as likely AI-generated may not be processed. CD Baby says it cannot accept any AI-generated content, even partly, and will remove it and possibly end the relationship. The three differ on whether the gate is a disclosure (DistroKid), a tool-provenance rule (TuneCore) or exclusion (CD Baby). Access note: the HTML pages returned a bot challenge to curl; the public Help Center ...
• articleWhat does Deezer do about AI-generated music (detect, tag, exclude, demonetise, remove), what can an uploader declare, and where do Deezer's own pages contradict each other?
  Across five Deezer press releases (2025-06-20, 2026-01-29, 2026-04-20, 2026-07-21, 2026-10-13), Deezer describes platform-side detection of fully AI-generated tracks (its own patent-pending tool, claimed to detect output of named generators), a listener-visible tag on albums that contain such tracks, automatic exclusion from algorithmic and editorial recommendations, demonetising of streams judged fraudulent, no longer storing hi-res versions (April 2026), a July 2026 takedown rule for fraud-linked tracks and tracks unstreamed for six months, and an announced listener toggle "before the end of 2026". No uploader-side AI disclosure field is described in these pages. All figures are Deezer's own. The July 2026 release dates the first tagging to June 2026 in one place and June 2025 in another, and keeps a notes-to-editors paragraph calling removal or demonetising "potential future actions" in the same release that announces takedowns.
• articleWhich DDEX ERN 4.3.2 elements and values carry an AI-generated flag for a release, recording or contributor, and how do YouTube, Spotify and DistroKid disclosures relate to them?
  In the ERN 4.3.2 XSD (Last-Modified 2025-06-30), a three-valued ContainsAI element (All / None / Partly, defined in the allowed-value-sets file) appears on Release, SoundRecording, Image, Text and Video, and a contributor-level AiContribution element appears on Contributor and EditionContributor, defaulting to All when SpecialContributor is GenerativeAI. Documented consistency rules: if a resource-level value is All only GenerativeAI contributors may be listed; if None no contributor may carry Partly or All. The release notes list recording-level and contribution-level AI communication under the Version 4.3.1 entry. YouTube accepts the declaration via DDEX ERN 4.3.2+ or a CSV column (all/partly/none/unknown) and reads a missing value as unknown. Spotify's provider page describes role-level credits for roles that were fully AI-generated, in a beta; DistroKid asks about lyrics, music, all of the audio or part of the audio. Which DDEX value a distributor writes for a given UI answer is...
• articleDo YouTube's public Content ID pages say anything about AI-generated music or video as reference material, and how does the separate music-partner AI declaration relate to Content ID?
  Negative result, scoped to the two pages read (S3 "Content eligible for Content ID" and S4 "Qualify for Content ID", captured 2026-10-22): neither contains the words AI, artificial, generative, synthetic or machine-generated (0 hits each after whitespace normalisation), so they state no AI-specific eligibility rule. Eligibility turns on exclusive rights, distinct individual references, and manual review for content sold or licensed at scale; open-licence and non-exclusive material is ineligible, as are sound effects and loops. YouTube's AI declaration for music partners (page S2) is a separate mechanism, delivered via DDEX or CSV, and that page does not mention Content ID (0 hits). Whether a given AI-generated track can satisfy "exclusive rights" depends on facts these pages do not address, such as the generator's output terms (see C3-B138-G49 for one service's terms). No legal conclusion is drawn. The Content ID partner agreement and CMS templates are not public and were not read.
• articleHow do robot world-model sources define action conditioning and what do they measure (video quality, value prediction, policy ranking), and which numbers support the claim that a world model can rank robot policies?
  Robot world models condition on low-level action vectors: a 7-D gripper delta on Bridge (Cosmos), whole-body command sequences (1XWM). Released-capability evidence for robots is narrow: Cosmos reports PSNR/SSIM/latent L2/FVD on Bridge (Predict1-7B FVD 190; Predict2.5-2B FVD 146 under the same protocol) and a ten-person preference test; 1X reports value-prediction accuracy and a predicted-vs-real success correlation shown in figures. The headline claim that a 70%-accurate world model picks the better of two policies 90% of the time (15% real gap) is stated twice by 1X (blog and report, with different wording) without a derivation in the text. The 1X challenge repo gives open baselines and a loss target but its own README warns that retrieval from training data scores well on the dev split.
• articleHow should an agent read FID/FVD tables on nuScenes for driving world models, given that one paper reports two sets of numbers for the same method and later papers copy different ones?
  The same method (DriveDreamer) has two published nuScenes results: FID 52.6 / FVD 452.0 in arXiv v1 and FID 14.9 / FVD 340.8 in v2, with the DriveGAN baseline also changing (73.4 / 502.3 to 27.8 / 390.8). Vista's Table 2 reuses the v1 pair; DriveDreamer-2's Table 1 uses the v2 pair. Vista says no prior model was publicly accessible, so cross-paper comparisons are of reported numbers, each under its own sampling, resolution and conditioning protocol. Downstream detection/tracking gains (mAP, NDS, AMOTA with generated training data) are an additional, task-based check used by DriveDreamer and DriveDreamer-2. A ranking claim from a table is only as good as the protocol match.
• articleHow do the driving and robot world-model sources measure multi-view, temporal, physics and control-following consistency, and which of those measurements are numbers versus figures or company claims?
  The sources use five families of checks: distribution distances (FID, FVD, FDD, FVMD), epipolar Sampson error within a view over time (TSE) and between views (CSE), pose re-estimation from the generated clip to check commanded paths (Cosmos TAE/TFE and camera-path errors, Vista IDM trajectory difference), detector/segmentation agreement with the conditioning boxes (GAIA-2), and human preference tests (Vista, Cosmos instruction tests). GAIA-2 provides no multi-camera metric in its text; Cosmos does (TSE/CSE, Table 24 and Predict2.5 Table 14) and notes the real-video reference values. Wayve's GAIA-3/4 pages describe LiDAR alignment, a three-level fidelity scheme and unpublished correlation claims; these are company statements without a linked report. Numbers from different reports are not comparable even for similar model names.
• articleHow far can I trust a MEt3R-style score that rests on a learned reconstruction backbone: what do controlled-corruption, injected-warp and human-ranking tests say, and what do COLMAP-based scores do better or worse?
  Three independent tests probe MEt3R from outside. SysCON3D (v1) finds that learned backbones (DUSt3R, MASt3R, VGGT, Fast3R) produce geometry for unrelated or noise inputs, MEt3R scores such sets as more consistent than real ones in its harness (negative effect sizes, 23% overall win rate), and its rank correlation with humans on eight sparse-view NVS methods is -0.10 to 0.26 against 0.98 for a COLMAP-based score. GeCo's injected-warp test finds MEt3R near chance at spotting a deformed frame. The SGC paper, in a different setup (text/video generators, 11 entities), reports MEt3R at Spearman 0.864 with humans. The sources agree that MEt3R is no stand-alone gate; they disagree on how well it ranks generators.
• articleWhat does WorldScore's DROID-SLAM reprojection-error '3D Consistency' score measure and miss, what do named closed and open video models score on it, and how much has it been validated?
  WorldScore's 3D Consistency runs DROID-SLAM on a generated static-world clip and averages the reprojection error of co-visible pixels between consecutive frames, then maps it to 0-100 against an empirical worst case (a frame-interpolated clip between unrelated images). Reported scores: Gen-3 68.31, Hailuo 67.18, DynamiCrafter 72.90, CogVideoX-I2V 86.21, 3D scene generators LucidDreamer 90.37 and WonderWorld 86.87. Its authors check it with human pairwise tests (91.7% and 97.3% agreement for 30-point bucket gaps); an independent injected-warp test found it good on scene-level but weak on object-level deformation. In code it assumes a fixed lens and cannot score dynamic subjects.

More