Evaluation and media provenance

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

• articleIf I put happy or sad music under a close-up of a neutral face, will viewers read the face differently, and what should I check first?
  One experiment tests exactly this: in Baranowski's study (30 students, thesis chapter and journal abstract), happy music made actors look happier and sad music less happy, with the effect strongest when the face itself was neutral; sad music did not make faces significantly sadder than no music. A different open study (Ansani et al. 2020) found music changed empathy and impressions of a character but not the overall attributed emotional state. No independent replication with a neutral face was found in the open sources checked. So music can be expected to push reading in its own direction, modestly, and the no-music version is the baseline to compare against. The sources do not say who owns the music used, so clearance is a separate step.
• articleWhich licence or terms text covers each of the five 3D asset sources behind TRELLIS-500K (Objaverse-XL, ABO, 3D-FUTURE, HSSD, Toys4k), and what does each text say about commercial use, redistribution and training?
  The five sources carry five different kinds of terms, and none of the five texts mentions model training. ABO is CC BY 4.0 as a whole. Objaverse-XL is ODC-By 1.0 as a whole with the licence of each object left to the object. 3D-FUTURE (agreement of March 28, 2020) is research-only with bars on commercialisation and on handing the data to third parties. HSSD is CC BY-NC 4.0 (project page) and gated; its paper says the Floorplanner data is made available for academic use. Toys4k has no licence file in the repository: it asks users not to distribute the data and refers to a separate attribution PDF that was not read. This is a reading of the texts, not a legal conclusion.
• articleDoes ending an episode on a cliffhanger make viewers watch more, according to open empirical studies?
  Two open-text studies disagree in outcome and differ in design. Wirz et al. (2022, N=133 adults, 3-4 episodes of drama) found cliffhangers raised arousal but did not raise enjoyment or intention to continue. Lu et al. (children, 12 serial vs episodic animation episodes in a game study) found serial plots sustained play and activity but not immersion or intention. Neither shows that a cliffhanger by itself binds viewers.
• articleWhat do the WER and speaker-similarity (SIM) numbers produced by the Seed-TTS-eval scripts actually compute, and what do they not tell me about a generated voice?
  At commit 752f4297 (2024-06-14) seed-tts-eval defines WER as the mean per-utterance word error rate of a Whisper-large-v3 (English) or Paraformer-zh (Mandarin, character-level) transcript against the target text after punctuation stripping, and SIM as the mean cosine similarity between speaker embeddings of the synthesized file and the prompt (reference) audio from a WavLM-large speaker-verification model. Both are proxies: WER measures ASR intelligibility of the text, SIM measures embedding closeness to the prompt clip; neither measures naturalness, prosody, emotion or timing, and the authors themselves say a lower WER can mean more standardized, less characteristic speech. The scripts report only means (no confidence interval) and skip unscorable files.
• articleOf the free sound-effects libraries I might use (Freesound, Sonniss GDC bundle, BBC Sound Effects, Pixabay sound effects), which say anything about AI or machine-learning training, in which document, and does the short summary page say the same as the legal text?
  A comparison at dated reads. Freesound: training allowed if each sound's CC licence is respected, plus uploader preferences since July 2026. Sonniss: AI training expressly prohibited in the licence and repeated on the bundle page. BBC: AI development or training needs BBC permission. Pixabay: the licence summary page says nothing about machine learning, while the Terms of Service prohibit scraping and data mining "for machine learning purposes" and describe a contributor opt-out for Pixabay's own AI training; so a reader of the summary alone would miss the ML clause. Mixkit's sound-effects licence text is script-rendered and was not obtained.
• articleWhat does the Sonniss #GameAudioGDC bundle licence (version 2.0, effective 27 August 2026) permit and forbid for commercial use, redistribution, attribution and AI training, how did it change, and where does the bundle page summary differ from the legal text?
  At the snapshot of 2026-10-10 the licence version 2.0 allows use and modification in personal and commercial projects without attribution, bars selling the sounds as they come and bars supplying them to others as sound effects (also in libraries or packs, modified or not), and expressly prohibits using them to train, develop or enhance AI technologies. The history page says the AI section was added on 28 January 2024, that earlier text allowed free distribution, and that the version published on the download day governs. The bundle page summarises the sounds as licensed for media production "only" and the AI ban as "training", both narrower or different in wording than the legal text.
• articleWhat scores do these papers report on HumanML3D-based benchmarks, how comparable are they, and what could an agent check before animation review?
  PhysDiff Table 1 reports on HumanML3D (mm): MDM Penetrate 11.291, Float 18.876, Skate 1.406, Phys-Err 31.572; PhysDiff w/ MDM 0.998, 2.601, 0.512, 4.111; MotionDiffuse 20.278, 6.450, 3.925, 30.652 (rows sum correctly; no mocap ground-truth row). GMD reports MDM foot-skating ratio 0.284 and 0.161-0.309 across variants in trajectory-conditioned tests. Scores come from authors' own runs on different protocols, so they show direction (physics projection reduces artefacts) not a portable threshold. HumanML3D is 20 fps, so a clip's frame rate matters for any per-frame threshold.
• articleHow could an agent test whether a line's subtext is recoverable, using the designs of the open studies, and which parts are measured vs my proposal?
  No inspected source gives a validated test for authored subtext. The open studies do offer three measurable designs an agent can adapt: forced-choice paraphrase of an ambiguous final line with a knowledge-state manipulation (Trott and Bergen), category classification plus psycholinguistic ratings of statements in context (Valles-Capetillo), and independent labelling with agreement reported (MUStARD). The protocol below is a synthesis built from those designs, not a result.
• articleWhat cues make an audience hear irony or sarcasm in a line (context mismatch, tone of voice, face), which dominates, and what does one TV-dialogue sarcasm dataset show about context?
  Two open-access experiments agree that mismatch between the utterance and its context is the main driver of irony/sarcasm recognition; prosody modulates it. In Spanish speakers (Valles-Capetillo 2022) contextual discrepancy gave the highest classification accuracy among three tested cues, though prosody and face also worked and were faster. In 22 Japanese listeners (Nakamura 2022) negative prosody raised sarcasm ratings in either context while positive prosody lowered them only in a positive context. A TV-show corpus (MUStARD, ACL 2019) shows ironic lines are hard to label reliably (kappa 0.15 initially) and that multimodal cues helped a classifier, but adding dialogue context helped little. Limits: stimuli are constructed scenarios or labelled clips, not authored scenes.
• articleWhich title-safe and action-safe numbers do EBU, SMPTE (via secondary extracts), DPP/Channel 4, Netflix, BBC and ATSC documents give for HD and UHD, and where do they disagree?
  Readable sources converge on a 90% title/text/graphics rectangle (5% inset per side) and a 93% action rectangle (3.5% inset): EBU R 95, SMPTE ST 2046-1 as quoted by NAB, Netflix (2026 snapshot), and the DPP/Channel 4 text-safe tables (90%) agree numerically. They differ in what they require: Netflix says it has no strict requirement and tolerates motion text leaving the box; DPP/Channel 4 make text safe mandatory and add a 4:3 text-safe box of 67.5% x 90%; ATSC maps ST 2046-1 onto IMSC1 subtitles as origin 5% 5%, extent 90% 90%. Older rules differ: BBC Technical Standards v1.3 (2002, third-party copy) gives caption safe 80% width and 90% height for 16:9 protected UK TX, and the CEA-708 caption window still uses the legacy 80% x 80% area per SMPTE. The ST 2046-1 text itself is paywalled; SMPTE reports in 2024 list an open project to add UHD safe areas.
• articleHow consistent are readers when they state a story's theme, what do openly readable empirical studies actually measure, and where does the evidence stop?
  Among the openly readable full texts read here, none reports how well independent readers agree on the theme of the same text. What they report is (a) coding reliability for researchers classifying responses as literal or interpretive (84% agreement, kappa .70; kappa .82-.95 for think-aloud categories; 90% for children's lessons), (b) an instructional effect on the share of interpretive thematic statements in one small classroom study, and (c) a developmental progression in extracting a fable's lesson. Two experts' statements for the same excerpt are related but not identical, and the authors say expert readings are not definitive. The one study aimed squarely at reader variance (Kurtz & Schober 2001) was not openly readable.
• articleWhat did the openly readable Seifert/Dyer-line experiments measure about readers recognising or connecting stories that share a thematic structure (thematic abstraction units), and what does that imply for relying on readers to infer a theme?
  The open Seifert, McKoon, Abelson and Ratcliff (1986) paper reports six experiments with short stories built on Dyer's thematic abstraction units (goal-plan patterns). Speeded verification showed no priming between thematically similar stories unless subjects were instructed to rate similarity or had prestudied the stories, so the authors conclude thematic connections are not made automatically. The paper also reports earlier findings (read second-hand) that people can sort and write stories by thematic pattern. Dyer 1983 and Seifert, Dyer and Black 1986 (Text) were not openly readable, so theme-abstraction itself is known here only through this paper's description.
• articleExactly how does the Matrix-Game GameWorld Score measure keyboard and mouse following, what do Oasis, MineWorld and Matrix-Game 1.0/2.0 score on it, and what does it need to run?
  GameWorld Score (Matrix-Game 1.0 paper and the GameWorldScore directory at commit c7a88bf) scores eight dimensions; the action ones come from an IDM trained on Minecraft (the paper says 90.6% keyboard accuracy, R2 0.97 for mouse) applied to generated clips. The code asks for 76 actions on 32 initial images, i.e. 2432 videos. Paper results (author-run): keyboard/mouse accuracy 0.77/0.56 Oasis, 0.86/0.64 MineWorld, 0.95/0.95 Matrix-Game 1.0. Matrix-Game 2.0 reuses the score on Minecraft (0.91/0.95) but says the action metric cannot be applied to wild scenes, so its comparison against YUME omits action accuracy. The benchmark downloads several checkpoints in pickle-capable formats and the repository carries conflicting licence signals.
• articleWhat evaluations of AI directors and emergent-narrative systems do the designers and open papers actually report, and how strong is the evidence that they improve player experience?
  The designer sources report outcomes, not measurements: Valve cites sales and awards for Left 4 Dead and asserts that a crude intensity estimate yields good pacing; no pacing data appear in the slides read, and RimWorld and Dwarf Fortress sources report none. A 2024 open paper says directly that no comparative data exist for the L4D director. Its own human-subject study of three directors in a small farming game (108 started, 32 usable) found that non-random directors beat random on time played and quests presented, with enjoyment lower only for random in the second session, yet its conclusion says players failed to perceive a difference. Valve's playtesting deck lists physiological arousal measures but contains no director-specific results. The evidence base for pacing controllers is therefore thin and mixed.
• articleCan a language model's next-word prediction error stand in for where human listeners perceive event boundaries in a story?
  Kumar et al. (2023) used GPT-2 XL to compute four per-word measures (surprisal, entropy, Bayesian surprise as KL divergence between successive next-word distributions, and embedding cosine distance) on three audio stories and tested them against human button-press segmentation with leave-one-story-out cross-validation. Only transient Bayesian surprise reliably predicted boundaries; correlations were small (about 0.05 to 0.12 in Table 1) though above a circular-shift null. This is a text-only audio-story study with an older model and no film, so it does not show that a language model finds scene boundaries in screenplays or video.
• articleUnder the Move One App EULA, who owns the mocap Output, what may I do with it commercially, and what may I not do with it (AI training, competing products)?
  At the Move One App EULA page fetched 2026-10-22 (header Last-Modified 2026-10-22T01:02:33Z), Move AI owns the IP in the Motion Data and Output (7.1.2) and licenses the Output back to the user worldwide, royalty-free, transferable and sublicensable (7.3), but the licence is void if Output is put to a defined Prohibited Use (7.5): building a competing product, training any AI System, or using it as an input to an AI System other than to create an original video work. The user in turn grants Move a perpetual, irrevocable licence to uploaded video for providing services and for developing and training its products (7.2). This is clause reading only, not legal advice or clearance.
• articleWhat do the Kuleshov papers state about where their face and context stimuli came from, whether the stimuli or data are public, and under what terms, so an agent knows what can be re-used?
  Articles are mostly CC BY (Cao, Calbi, Mullennix, Szaszko & Loebus), but the stimuli are a different matter. Face sets named in the papers are KDEF (Calbi/Barratt, Mullennix) and NIM-STIM (Mobbs); context sets are IAPS (Mobbs, Mullennix) and OASIS (Szaszko & Loebus); Cao and Szaszko & Loebus filmed their own faces and Cao shot its scenes with actors who consented to publication. The one stimulus-set terms text read (KDEF's own page) allows only researchers, non-commercial use and forbids redistribution without written consent. No paper states a licence for its own clips; Cao's GitHub repository (no licence file, tree lists only spreadsheets and statistical maps) does not contain the video stimuli; Mullennix's data are on Dryad (metadata says CC0); Szaszko & Loebus provide data on request. Terms for IAPS, NIM-STIM and OASIS were not read.
• articleWhat does openly readable empirical work show about viewers' comprehension, load and enjoyment of flashbacks in film, and how strong is it?
  A CogSci 2011 paper (Sevenants & d'Ydewalle) compared a 14-minute student thriller with flashbacks against a re-edited chronological version: 30 students rated characters and ranked scenes; 8 students had pupil size recorded. Story order was easier to reconstruct (r = +.868 vs +.749), characters' acts more salient, and pupils larger in 5 of 7 flashback scenes, but aesthetic ratings did not differ. A 2025 open-access paper (Djonov & Tseng; abstract read, full text blocked) reports 28 children 7-10 on two Disney flashback scenes, with temporal comprehension supporting causal interpretation. Evidence is small-sample and mixed.
• articleCan I use sound generated by MMAudio, HunyuanVideo-Foley or ThinkSound in a commercial delivery, according to what their licences and cards actually say?
  The texts differ per component. MMAudio: code MIT; checkpoints CC BY-NC 4.0 (non-commercial) with a README warning that training-dataset licences apply and commercial suitability is not guaranteed. HunyuanVideo-Foley: Tencent Hunyuan Community License, which says Tencent claims no rights in Outputs but excludes EU/UK/South Korea, bars using Outputs to improve other AI models, restricts use outside the Territory and adds a 100M monthly-active-user condition. ThinkSound: HF card says Apache-2.0 while the GitHub README says commercial use is not permitted (see -05). No text read grants clearance; this record does not conclude legal permission.
• articleHow do I read the licence of a single Wikimedia Commons file safely: which part of the file page and API is the evidence, and what can the API flatten?
  Commons policy requires every file description page to carry a copyright tag, source and author, and says the page should allow others to verify the licence status; the help page for reusers adds that the Foundation gives no warranty and the reuser should verify. In the worked example File:Hopetoun_falls.jpg, the page wikitext carries {{self|cc-by-sa-3.0|GFDL|attribution=...}} (two licences), while the API's License and LicenseShortName fields show only CC BY-SA 3.0; GFDL appears in Categories and the Permission text. So the wikitext at a page revision is the better evidence, and the API fields are a convenient summary.
• articleWhat do V-JEPA 2 and V-JEPA 2-AC checkpoints come with in code, weights and outputs terms, and what does the public text leave open?
  At commit 204698b45b3712590f06245fbfba32d3be539812 the README says 'the majority of V-JEPA 2 is licensed under MIT' with three Apache-2.0 source files and says nothing about checkpoints. Weights are hosted two ways: dl.fbaipublicfiles.com URLs listed in the README (including V-JEPA 2-AC, a 11.76 GB object; HEAD only, not downloaded) and 8 Hugging Face repos whose card front-matter says 'mit' for six and 'apache-2.0' for the two ViT-g fpc64 pretrained cards; the cards contain no licence prose and the HF repos list no LICENSE file. V-JEPA 2-AC is not in the HF collection. The paper (arXiv v1) says it trained V-JEPA 2-AC on under 62 hours of Droid robot video and pretrained on public datasets including YT-Temporal-1B; rights of that data are not addressed in the texts read. Terms for outputs, redistribution and training use of the weights: not stated.
• articleIs WHAM's context 10 seconds or 1 second, and do the two statements count different units?
  They count different things, and the card contradicts itself. The Nature paper's methods say the 1.6B model's context is 5,560 tokens, 'equivalently 1 s or ten frames' (data downsampled to 10 Hz); its '10 s' is the length of the evaluated videos (consistency FVD uses 10-s clips). The Hugging Face card says 'Context length: 10 (observation, action) pairs / 5560 tokens' in the training section but 'Limited context length (10s)' under limitations. Ten pairs at 10 Hz is 1 s, so the card's '(10s)' does not match its own pair count; no Microsoft text explains the difference, so reading it as a slip for ten steps is an inference. Code defaults agree with ten frames: run_dreaming.py sets context_length 10.
• articleHas the benign-violation theory been tested on humor that is not a moral-violation vignette, for example complaints and reviews, and with what methods and sample sizes?
  McGraw, Warren and Kan (JCR 2015) apply the benign-violation theory to humorous complaints in six studies using Yelp data (330,071 reviews), Facebook status updates, a Haribo review sample and mTurk experiments (one with N = 1,214). The reported pattern is that humor makes complaints read as more positive but praise as more negative, and that humorous complaints are less likely to win redress or sympathy. This is an extension by two of the theory's originators with non-moral stimuli, not an independent replication, and the open copy found carries a JSTOR-download footer, so its reuse status is unclear.
• articleWhat is the Kuleshov effect, and what do the open-access experiments say an editor can and cannot rely on?
  The 1920s footage is lost and the contemporary accounts conflict, so the classic claim rests on modern experiments. In the five open-access experiments read (Mobbs 2006, Calbi 2017, Mullennix 2019, Cao 2024, Szaszko & Loebus 2026) an emotional context shot shifted ratings of an unchanged neutral face, but the pattern is narrower than the folklore: valence shifts are the most repeated result, negative or fear contexts shift ratings more reliably than happy ones, forced emotion-label results are mixed, and the most tightly controlled recent study found only tiny valence shifts and no label shift. An editor can treat context as a modest, direction-of-valence bias to be tested with viewers, not as a guarantee that viewers will name a particular emotion. Barratt et al. 2016 (the best-known replication) is paywalled; only its abstract was read.
• articleWhat structural statistics can I get from Twine, ink and Yarn Spinner without writing my own parser, and what do they not tell me?
  At the pinned versions: Twine's Details view reports characters, words, passages, links and broken links; inklecate's -s flag prints words, knots, stitches, functions, choices, gathers and diverts as JSON (source-level, not for compiled .json); `ysc graph` exports all nodes and links as DOT or mermaid. None reports number of endings, path counts, cycles or reconvergence; those need the exported graph plus your own computation.
• articleWhat do the PerseusDL canonical-greekLit repository, the Perseus Digital Library site, the Poetics TEI header and Harvard/Loeb public pages say about rights in the Fyfe (Loeb) English Poetics that Perseus labels CC BY-SA 4.0?
  No inspected public page is a Perseus or Harvard statement specifically about rights in Loeb-derived translations such as Fyfe's Poetics. What exists is a layered set of generic statements that disagree on the licence version: the TEI header and repository README apply CC BY-SA 4.0 ('unless otherwise indicated'), the Perseus site page for the same text says CC BY-SA 3.0 US, and the Perseus copyright page says Tufts holds overall copyright, rights vary per component, and some texts (Loeb/Harvard University Press named) are used 'by agreement with the publishers who hold their copyrights'. The Poetics text itself carries no copyright or public-domain assertion; its only rights-relevant metadata are the licence element and a Heinemann/Harvard sourceDesc. The question of what Harvard permits was not answered by any inspected Harvard/Loeb page.
• articleCan an L10 experiment use Xiph's derf test-video collection with a clear per-clip licence, and what do the per-clip notices say?
  Not as a collection. The derf page says clips were publicly accessible or given to Xiph and are 'believed to be freely redistributable', that some may impose additional restrictions, and points to copyright or readme files. By my regex tally of the HTML, 118 clip entries, 85 with a copyright or readme link in the header and 32 with no link; 9 entries say login required at the original source. The notices I read differ: Fujitsu's Bowing notice grants use only for research purpose with conditions, the Hannover-path notice limits use to developing, testing and promulgating technology standards, and the twitch folder says public domain with a fair-use caveat. So derf clips need a per-clip read.
• articleWhat exactly does Microsoft's WHAM/Muse release offer for gameplay ideation, and how do the Nature paper and the model card differ?
  WHAM is a single-game model (Bleeding Edge), 300x180 at 10 Hz, with a 1-second context window, built from a user study of game creatives and evaluated on consistency (FVD), diversity (Wasserstein) and persistency (blinded human labels). The released 200M and 1.6B checkpoints are research-licensed (non-commercial, no distribution), not real time, and domain-bound. The model card's 'Limited context length (10s)' conflicts with the paper's 1 s of context; the card's own '10 pairs' agrees with the paper.
• articleI received a third-party media, model, dataset or code asset: what do I need to find out and record about it before I can credit it, and which text tells me what is required?
  A hosting platform's licence tag is a pointer: Hugging Face model-card metadata takes a licence identifier or 'other' with a name and link to a custom licence, so the licence prose is what must be read. Licences can split by component: the CreativeML OpenRAIL-M text at the pinned commit states the Data is not licensed under it. REUSE's .license sidecars can carry per-file licence data for media only by analogy, since REUSE scopes itself to software. Proposal (producer synthesis, labelled): order of work identify exact object and revision -> read licence prose per component at that revision -> split rights by component -> record provenance -> assemble attribution. Not legal advice or clearance; each asset's own licence text must be read.
• articleWhat do the C2PA conformance programme, certificate policy, trust list and conforming-products list require before a signer's manifests validate as trusted, and what route do those texts imply for an agent platform that wants to sign output?
  Under C2PA's v0.2 conformance documents, a manifest is 'trusted' in the C2PA sense when its claim-signing certificate comes from a Trust List CA. Such certificates are issued only to instances of a conformant Generator Product listed on the Conforming Products List, with conformance proven by a signed Notice (a public CPL record may also serve). The private key must stay under that instance's exclusive control. On 2026-10-22 the Trust List (issued 2026-08-05) held 18 entities, and the CPL held 270 conformant records: 240 generators (229 at assurance level 1, 11 at level 2) and 30 validators. Nothing in these texts addresses agent platforms specifically.
• articleCan I treat the c2pa-rs test images and clips (earth_apollo17.jpg, mars.webp, Big Buck Bunny clip, sample1.*) as covered by the repository's MIT/Apache licence?
  The c2pa-rs licences cover the crate's software (MIT and Apache-2.0, dual) and the README warns that some components carry different terms. Nothing beside the test fixtures states their origin or terms, the published crate excludes tests/fixtures, and the project's own contributions note calls third-party test images something that need not be in the SDK while wishing for public-domain replacements. Treat the images and clips as unlicensed for reuse until their origin is established elsewhere.
• articleWhat is the AOUSD USD Core Specification, which document is the published version, and where do I read it (and under what licence)?
  The only published AOUSD Core release in the repository is v1.0.1 (2025-12-12, PDF and Markdown), which AOUSD's announcement calls Core Specification 1.0. It is CC BY-ND 4.0: verbatim copies with attribution may be redistributed, but distributed adaptations may not.
• articleBeyond SFU, ACCAD, CMU and KIT, which further AMASS component datasets state their own licence on public pages, and what do they say?
  Three AMASS-related components state their own terms: HDM05's page (archived 2021, last modified 2015) labels the database CC BY-SA 3.0; BMLmovi licenses its data for non-commercial research only, prohibits training methods or neural networks for commercial use, and says the part included in AMASS falls under the AMASS licence; Dynamic FAUST's MPG licence covers 'Software/Data' for non-commercial research and also prohibits training neural networks for commercial use.
• articleWhat does the Wikimedia Commons imageinfo extmetadata return for licence and credit, and how does the extension documentation qualify those fields?
  Wikimedia's CommonsMetadata extension describes its extmetadata fields as a temporary solution that outputs parsed HTML and warns that licence values are unreliable for multi-licensed images (with License a best guess and LicenseShortName probably more reliable); a sampled file (Hopetoun_falls.jpg) reports CC BY-SA 3.0 in the licence fields while its categories include GFDL and 'License migration redundant', i.e. a multi-licence case the single licence field does not show.
• articleWhat ExifTool sample files exist for metadata testing and what do the repository and sample-image page say about their reuse?
  At ExifTool 2200871d the README says it is distributed on the same terms as Perl (Artistic License or GPL), while the LICENSE file is the GPL v3 text; the repository ships 194 verification test files under t/images (722,199 bytes) with no file-level licence statement found; the separate sample-image archive replaces main images and strips previews 'to avoid copyright problems' while preserving EXIF/APP segments, without stating a licence.
• articleWhich licence applies to Netflix Open Content titles in AV1/, av2ctc/ and IMF/, and does it match what the Open Content page says?
  Netflix's Open Content page says its open source content is available under CC BY 4.0, but the per-title licence files for the Chimera AV1 encodes (Sept 2019) and the Meridian IMP/VDM/Atmos assets (Nov 2016) license those sequences under CC BY-NC-ND 4.0, and the AV1 readme says each asset's licence is in its own license.txt; the TechblogAssets Sparks licence file referenced by the README returns 404.
• articleWhat do the LRS2 and LRS3 pages and licence file say about permitted use of the data, and do they address model training?
  LRS2 is offered for non-commercial academic research only after signing a BBC R&D Data Sharing agreement (agreement not read); LRS3 is described as available for research purposes under CC BY 4.0 while its licence file says the underlying TED video excerpts must respect TED terms and CC BY-NC-ND 4.0, copyright stays with the video owners, and downloads are no longer offered on the site. None of these texts mention model training.
• articleWhat terms do MedleyDB and FMA state for their audio, annotations and metadata, and do they address model training?
  MedleyDB's pages offer the dataset for non-commercial research under Creative Commons (named both as Attribution-NonCommercial and as BY-NC-SA 4.0) and ask users not to republish without consent; the Zenodo audio record (v1.0) is access-restricted with no licence field and points to a GitHub repository for annotations whose LICENSE is MIT, while the description page says audio and annotations are released under BY-NC-SA. FMA separates MIT code, CC BY 4.0 metadata and per-artist audio licences, for research purposes. Neither text addresses model training.
• articleWhat is the status of APCA and WCAG 3 contrast relative to WCAG 2.2, what does the WCAG 3 draft actually require, and where do APCA and WCAG 2 differ?
  The WCAG 3.0 Working Draft of 10 September 2026 leaves its text-contrast measure 'to be determined' and does not mention APCA (curator search of the same bytes the producer pinned found no occurrence); apca-w3 (6c5c066c) uses a 2.4 exponent and polarity-specific constants, structurally unlike WCAG 2's symmetric ratio, and both APCA repositories state patents pending and restrict use (apca-w3: WCAG web content via a W3/AGWG agreement; SAPC-APCA: registered beta testers or personal use unless authorised in writing).
• articleWhich AMASS component datasets state their own licence or terms on their own public pages, and how do those statements differ from AMASS's website and repository licence?
  The AMASS website licence treats the AMASS data and software as one Dataset licensed only for non-commercial research, education or artistic projects, while component sources state their own differing terms: SFU says research-only and no commercial products or resale, ACCAD labels its Open Motion Project CC BY 3.0 Unported, and CMU's info page says its motions are free to download and use. The pages read do not say how component terms apply to AMASS's fits of the same captures.
• articleWhich glTF-Sample-Assets models carry non-CC0/CC-BY terms (non-commercial, EULA, testing-only) that L10 must exclude or handle separately?
  At 15cd02d3 the repository's REUSE.toml has 151 annotation blocks: 62 plain CC0-1.0 and 46 plain CC-BY-4.0 (curator re-tally matches the producer); others combine CC licences with trademark references or name non-open terms (SCEA, CRYENGINE agreement, 3DRT testing-only, Poser EULA, Adobe Stock, CC-BY-NC-SA-4.0, CC-BY-4.0 AND CC-BY-NC-4.0). VirtualCity's 3DRT grant is for testing glTF tools only, and DamagedHelmet lists CC-BY-4.0 and CC-BY-NC-4.0 together without file-level scope.
• articleWhat consent, watermark and disclosure mechanisms does VibeVoice document, and are they still enforced in the repository?
  The VibeVoice-1.5B HF card (c00898d2, MIT front matter) documents consent-based out-of-scope uses, an audible disclaimer and an imperceptible watermark; the GitHub README (16fb2cb1) says the TTS code was removed on 2025-09-05 after misuse and marks the 1.5B TTS entry Disabled. Card and README describe different artefacts at different revisions.
• articleFor keeping one person's face identity across storyboard panels, what do the PhotoMaker and InstantID papers and official repositories measure, what limits and compute do they state, and how are their code and checkpoints licensed?
  PhotoMaker (arXiv 2312.04461v1; repo 060b4fc) stacks 1-4 ID photos into one embedding for SDXL; its Table 1 reports PhotoMaker CLIP-I 73.6, DINO 51.5, Face Sim. 61.8 at 10 s on a V100 (IPAdapter has higher Face Sim. 67.1), trained on 8 A100 for two weeks; it handles one person per image and is better at half-length portraits. InstantID (arXiv 2401.07519v2; repo 2145b67) uses one face photo with an InsightFace embedding, a ControlNet-style IdentityNet and an image adapter, trained on 48 H800; the paper has no quantitative table (qualitative comparisons) and the README supports only the largest face. Licence components differ: both codebases are Apache-2.0 (PhotoMaker's LICENSE header mentions third-party components but lists none after the Apache text); InstantID's README says InsightFace face models are non-commercial research only and its released checkpoints are research only, while the Hugging Face card says the project is Apache.
• articleHow does ConsiStory keep one subject consistent across prompts without training, how was that measured, what are its stated failure modes and costs, and can the official code be used commercially?
  ConsiStory (arXiv 2402.03286v3, SIGGRAPH 2024 TOG) shares internal activations between the images of a set so that a subject stays consistent without any optimisation step; the first two prompts act as anchors (paper: two anchor images, 0.5 dropout). The authors evaluate 100 image sets with DreamSim (background removed) and CLIP and a 3,000-response two-alternative user study; the quantitative headline is a plot, not text. Timing: 32 s on an H100 for two anchors plus one new image, about 25x faster than The Chosen One (their estimate of 13 min). Stated limits: it relies on cross-attention localisation (failures below 5% in their observation, fixable by changing the seed), images in a set must share one style, and SDXL biases carry over. The official NVlabs code is under the NVIDIA License, which permits only non-commercial use and only with NVIDIA Processors.
• articleWhich of the multi-shot and identity-consistency systems and benchmarks in these papers are actually released, as what (code, weights, data), and under which stated terms?
  Released: HoloCine (14B full and sparse weights as .safetensors on Hugging Face, code, CC BY-NC-SA 4.0), StoryMem (code, Hugging Face weights, S-Lab licence non-commercial; Hugging Face metadata says CC BY-NC 4.0), and OpenS2V-Eval (code Apache-2.0). Partly released: MuSS (data-construction code and YouTube ID lists only). Not found released: EntityBench data/code, EntityMem, PersonaShot, MuSS checkpoints. Both HoloCine and StoryMem require separately downloaded Wan2.x base weights (HoloCine README names .pth files, a pickle-capable format) whose terms were not read here. None of the licences were interpreted for output use.
• articleWhere do licence statements, dependency notices, dataset provenance and weight file formats disagree or leave gaps for these 3D generators?
  Four items need curator attention: (1) TripoSG advertises MIT but its NOTICE lists code derived from Tencent HunyuanDiT (Tencent Hunyuan Community License notice) and the BRIA RMBG-1.4 background-remover whose card says non-commercial; the README downloads RMBG weights. (2) TRELLIS-500K metadata says MIT while the DATASET.md describes assets curated from Objaverse(XL), ABO, 3D-FUTURE, HSSD and Toys4k and states the dataset is CSV metadata; asset-level licences were not read. (3) Hunyuan3D repos ship .ckpt (pickle-capable) next to safetensors; 2.1 ships .ckpt and .bin; TripoSR ships model.ckpt; the licence texts do not address file formats. (4) The Tencent Hunyuan 2.0 HF card links its licence at LICENSE.txt while the HF file list shows LICENSE.
• articleWhich systems claim to steer a running world model or long video with text, how was that steering actually tested, and what did the tests show?
  Four measured protocols for text steering exist in 2025-2026 preprints, plus one system (Zing-0.5) that shows it only qualitatively, and each measures something different. Hunyuan-GameCraft-2 uses a VLM judge on six interaction dimensions over 100 test images against image-to-video baselines. Incantation uses blinded human ratings (three annotators, 400 clips) of whether an entity performed the prompted action. LongLive uses CLIP score per 10 s segment over 160 videos with six prompts each. WorldSonus compares a mid-stream sound prompt switch against a no-switch hold with identical visuals (audio only). Zing-0.5 measures only navigation (WBench) and shows text events as a qualitative session whose lasting consequences the authors say are not established. No source tests a text event in a released interactive model with persistence over later actions.
• articleFor the audio-capable world models and sync-focused sound tools found here, what do the repositories and model cards actually say about licence and weights per component, and where do they disagree?
  HelixWorld (Noiz AI) is the one audio-visual world model with a published checkpoint: code and the HF card say Apache-2.0, but the repository NOTICE says the LTX-2 code and model assets inside it are governed by the LTX-2 Community License (paid licence for entities with at least US$10M annual revenue; fine-tuned weights count as derivatives), the paper says HelixWorld builds on LTX-2.3, and the Gemma text encoder has its own terms. WorldSonus is CC BY-NC 4.0 with .pt weights (pickle-capable). V-AURA code is MIT but its checkpoint is a tar.gz with no stated weight terms. Kling-Foley publishes a paper and a CC-BY-4.0-labelled benchmark dataset but no code or weights were found.
• articleWhat do Freesound's per-sound Creative Commons licences, attribution guidance and AI-training answers say at the access date, and where does Freesound's own summary go beyond or differ from the Creative Commons legal texts?
  Freesound distributes each sound under the licence its uploader chose (CC0, CC BY or CC BY-NC 4.0, plus legacy Sampling+). Its FAQ now answers AI training as "yes as long as individual sounds' Creative Commons licenses are respected" and, as of July 2026, adds uploader-level Gen AI preferences (open-source-only, non-commercial open-source-only, no generative AI, and a voice opt-out) on top of the licence, exposed to API users for filtering and in Data Packs. Freesound labels its CC-to-training reading as its own interpretation (for example CC-BY requires disclosing the training set); the CC legal code text does not mention AI and the CC FAQ says the answer depends.
• articleFor a Sketchfab model, which document governs what I may do (the per-model Creative Commons label, the Standard or Editorial licence, or the Terms of Use), and how do the NoAI and CreatedWithAI labels bear on use in a production or an AI dataset?
  Sketchfab layers three documents: a per-model licence label exposed by the API (seven CC types plus free-st, st, ed), the paid Standard/Editorial licence agreement, and Terms of Use effective 12 August 2026 that bar using NoAI-tagged content in generative-AI datasets and say the licence in force at download governs. Sketchfab's own 2023 blog says the NoAI tag may not bind outside the platform for CC-licensed models. The API objects inspected show no NoAI field, and the licence agreement page text read has no NoAI clause although the blog says Standard-licensed downloads are contractually covered; these gaps are for curation.

More