Evaluation and media provenance

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

• articleWhich cited benchmark papers contain the SGC scores for Cosmos and Latte, and what do the cited papers actually report for Sora 2 and Veo 3.1?
  Correction (unsupported by the cited sources). C3-B139-W28-07 reports 'SGC scores Cosmos 0.0722 and Latte 0.3226'. Searching the four cited papers (MEt3R 2501.06336, WorldScore 2504.00983, 2511.18991, GeCo 2512.22274) for 0.0722 and 0.3226 found neither value, and none of them evaluates Cosmos or Latte under an SGC score in the text extracted. The other numbers were reproduced: GeCo-Eval evaluates Sora 2 and Veo 3.1 (object-centric fused 0.069 and 0.090), the MEt3R paper lists 0.032, and WorldScore's 3D Consist column gives 68.31 for Gen-3 and 67.18 for Hailuo. Remove or re-source the SGC figures. 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.
• 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.
• 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.
• articleC3-B511-02 relayed a creator's Nov 2025 remark that Weavy "just got acquired by Figma". Figma's own site now lists "Figma Weave" in its newsroom navigation as "AI workflows for imagery, video, audio, and more" with a product domain weave.figma.com (R1), and a solutions page states "Every major...
  Corrected 2026-10-23 - see C3-B613-04; the original statements below are kept for history and are superseded where the correction says so. C3-B511-02 relayed a creator's Nov 2025 remark that Weavy "just got acquired by Figma". Figma's own site now lists "Figma Weave" in its newsroom navigation as "AI workflows for imagery, video, audio, and more" with a product domain weave.figma.com (R1), and a solutions page states "Every major AI model, in one place" with image, video, audio and 3D models from Google, Kling, Runway, Luma and others, selectable per workflow step "right from your canvas" (R2). Secondary reports located by search (R3, not read in full) place the announcement around 1 Nov 2025, describe Weave as the new product line built from Weavy, ~20 staff joining, and price estimates of $150-200M (Globes, market sources) - unverified. For agents: name the product "Figma Weave (formerly Weavy)" in current answers; Weavy-era tutorials (model import from fal/Civitai/Replicate, comp...
• articleEBU R98 structures BWF CodingHistory rows with keys A=, F=, B=, W=, M= and a free-text T= that should contain no commas; tool name, version and arguments therefore have no dedicated key and must be carried in T= by...
  EBU R98 structures BWF CodingHistory rows with keys A=, F=, B=, W=, M= and a free-text T= that should contain no commas; tool name, version and arguments therefore have no dedicated key and must be carried in T= by convention.
• articleTurning your own recorded sample into a full track (Suno "remix and cover") and why that differs from remixing others' outputs
  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. Mikkel Lassalle (2026-06-30): record a short phrase (a Spanish-style guitar riff; any instrument or voice, even a phone recording), upload it in Suno (Audio > upload, identify audio content), then use "Remix and cover" with a short style prompt ("Spanish guitar", "acoustic ballad"); the result keeps the sample's character ("same DNA"). Rights reading with Suno's current terms (C3-B411-01): uploading material you own is a "Submission" (you keep ownership but grant Suno a broad licence); commercial use of the generated track requires a paid tier and a permitted download; the terms' non-commercial joint-ownership rule applies to "Remixes" made via features that let OTHER u...
• articleBWF MetaEdit (5b7f3d20) validation help says bext Version must be 0 or 1, while its bext parser reads the BWF v2 loudness fields when Version >= 2; the help text lags the...
  BWF MetaEdit (5b7f3d20) validation help says bext Version must be 0 or 1, while its bext parser reads the BWF v2 loudness fields when Version >= 2; the help text lags the code.
• articleWhat free AI video tiers actually give you (Jan 2026 creator test) and why a "backed by data" title needs checking
  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 (2026-01-24) tests three no-cost options: a no-login generator with separate action and camera prompt fields (camera field noticeably improves motion; free tier limited to its basic model), Grok Imagine (continuous variations as you scroll, edit-by-instruction, presets incl. an adult "spicy" mode, text-to-video), and Meta AI create (variety/weirdness/stylization sliders - stylization helps avoid the glossy 3D look - restyle, animate, lip sync). Shared free-tier limits he observed: queue waits of 5-10 min for a 5 s clip, weak prompt adherence needing rerolls, older/lower-quality models with morphing faces; he attributes this to compute cost. Packaging ...
• articleCreator head-to-head (Aug 2026): Seedance 2.5 vs Seedance 2.0 vs MiniMax H3 vs Flux 3 video on camera moves, acting, dialogue, music and multi-reference
  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. Dan Kieft (OpenArt link, 2026-08-21), each prompt run up to 3 times, best shown, 15 s clips. Specs as he states them on OpenArt then: Seedance 2.0 up to 15 s, 480p-4K (3,000 credits 1080p/15 s, 6,000 4K); Seedance 2.5 up to 30 s, 720p only at the time, up to 50 references (30 images, 10 videos, 10 audio, 30 s total), 1,955 credits/15 s; MiniMax H3 up to 15 s, 2K, 15 references, 1,350 credits/15 s, runnable locally; Flux 3 up to 20 s, 1080p, ~3,000 credits/15 s. His verdicts: complex camera moves (crash zoom, whip pan, earth zoom-out) - Seedance 2.5 best, Flux 3 worst; simple cinematic shot - Seedance 2.0 4K looked best; complex two-angle dialogue shot - Seedance 2.5 bes...
• articleFrom Mirelo AI GmbH terms (R1): Free Plan users may use the services and outputs only for non-commercial purposes, must visibly attribute Mirelo as sound provider, and outputs may carry a watermark (s2, s7.5); Paid Plan users may use them commercially within the scope of the chosen plan (s2,...
  From Mirelo AI GmbH terms (R1): Free Plan users may use the services and outputs only for non-commercial purposes, must visibly attribute Mirelo as sound provider, and outputs may carry a watermark (s2, s7.5); Paid Plan users may use them commercially within the scope of the chosen plan (s2, s5.2). s7.3: any IP that arises in outputs vests exclusively in the user, but "certain Output may not qualify for intellectual property protection" and use remains subject to plan limits and the Acceptable Use Policy, including restrictions on using outputs to train AI models. s7.4 Training License: the user grants Mirelo a worldwide, royalty-free, sublicensable licence to use Input and Output to train its models and improve services. Paid users can opt out for a personal account via a Studio setting; that setting "does not control organization API keys" - organisations must email legal to confirm or change the arrangement "before submitting content"; opt-out stops future additions and "new thir...
• articleOscars rules for films released in 2026 (99th Academy Awards): acting must be "demonstrably performed by humans with their consent", screenplays human-authored, AI use may be investigated - AI in production is not banned outright
  Reported by NPR (2 May 2026) from the Academy's announcement; the official rules PDF was not retrievable (403). Changes for the 99th Oscars: only roles "demonstrably performed by humans with their consent" are eligible for acting awards (excluding synthetic performers such as the AI "actress" Tilly Norwood); screenplays must be "human-authored"; the Academy "reserved the right to investigate the use of generative AI in any submission". The Academy framed AI like earlier technologies (sound, colour, CGI) whose rules evolve; using AI somewhere in production does not by itself disqualify a film (per commentary). Other changes: performers can receive multiple nominations in one acting category; International Feature eligibility widened (festival winners; not limited to one film per country). Practical for AI filmmakers aiming at awards/festivals: keep authorship records (who wrote, who performed, consent forms), expect disclosure questions about generative AI, and keep synthetic perform...
• articleClaims (R6, R7): Sonilo says its music models are "trained only on professionally licensed music, including the Shutterstock catalog", citing a Sonilo-issued 11 May 2026 press release and a Shutterstock LinkedIn post; its SFX training sources are separately described only as "compliant" (not the...
  Claims (R6, R7): Sonilo says its music models are "trained only on professionally licensed music, including the Shutterstock catalog", citing a Sonilo-issued 11 May 2026 press release and a Shutterstock LinkedIn post; its SFX training sources are separately described only as "compliant" (not the same catalogue). R6 itself notes that "Training-source rights and the permission to use generated output are separate." Terms (R5, s6.2): commercial use depends on the plan when the output was generated, and upgrading later does not extend rights to earlier outputs. Free: personal/non-commercial with "Music by Sonilo" credit; Creator (paid): still non-commercial; Pro: commercial including advertising, client work, monetised channels, apps, websites and indie games (under US$200k combined revenue and funding in the preceding 12 months); Premium: adds film, television, broadcast and studio games; Enterprise/API per separate agreement - so the fal listing's "commercial-use-safe" wording does no...
• articleFrom YouTube Privacy Guidelines (R1): "altered" or "synthetic" content is content that looks or sounds like a person but was significantly edited or generated by AI; examples are face replacement and simulated audio of someone saying things they did not say. To qualify for removal the content...
  From YouTube Privacy Guidelines (R1): "altered" or "synthetic" content is content that looks or sounds like a person but was significantly edited or generated by AI; examples are face replacement and simulated audio of someone saying things they did not say. To qualify for removal the content should show a "realistic altered or synthetic version" of the person's likeness; review considers whether the content is altered/synthetic, whether it is disclosed to viewers as such, whether the person is uniquely identifiable, whether it is realistic, whether it has parody, satire or other public-interest value, and whether it features a public figure or well-known individual "engaging in a sensitive behavior, such as criminal activity, violence, or endorsing a product or political candidate". General rules: the person must be uniquely identifiable (image, voice, full name or other identifiers - a first name or a fleeting image usually is not enough); complaints must be first-party except for...
• articleVerification of the creator-reported claim in C3-B512-01: the MPA press page (R1) is dated 12 February 2026 and quotes Chairman and CEO Charles Rivkin on the launch of "Chinese AI service Seedance 2.0": unauthorized use of U.S. copyrighted works "on a massive scale" "in a single day", operating...
  Verification of the creator-reported claim in C3-B512-01: the MPA press page (R1) is dated 12 February 2026 and quotes Chairman and CEO Charles Rivkin on the launch of "Chinese AI service Seedance 2.0": unauthorized use of U.S. copyrighted works "on a massive scale" "in a single day", operating "without meaningful safeguards against infringement", and a demand that ByteDance "immediately cease its infringing activity". So C3-B512-01's statement that the MPA demanded ByteDance cease infringing activity is confirmed; its "24 to 48 hours with no guardrails" window and the expectation of later restrictions remain creator characterisations. Search results (R2, not read in full) also describe later studio cease-and-desist letters (reportedly Disney and Paramount Skydance) and an MPA letter, and ByteDance saying it was strengthening safeguards - unverified here. Current ByteDance controls for Seedance 2.x (real-face restrictions, authorised portrait assets) are in C3-B436-01 and C3-B466-01...
• articleYouTube rules (R1): chapters split a video into sections with individual previews and may appear in transcripts. Manual chapters: in the description list timestamps and titles, first must be 00:00, at least three in ascending order, minimum chapter length 10 seconds; manual chapters override...
  YouTube rules (R1): chapters split a video into sections with individual previews and may appear in transcripts. Manual chapters: in the description list timestamps and titles, first must be 00:00, at least three in ascending order, minimum chapter length 10 seconds; manual chapters override automatic ones. Automatic chapters are on by default for new uploads ("Allow automatic chapters", per video, in bulk, or via Upload defaults > Advanced settings) but not all videos are eligible; chapters are unavailable if the channel has active strikes or the content may be inappropriate for some viewers; channels without access must apply for Advanced features. Google Search (R2): key moments let searchers jump to segments; Google tries to detect segments automatically but "will prioritize key moments set by you, either through structured data or the YouTube description". For YouTube-hosted video, the description timestamps are the mechanism (all Search languages). For self-hosted or embedded ...
• articlePost-shoot AI controls on fal H3 Max (Oct 2026): keyframed orbit camera from one still (azimuth/elevation/distance on normalised time), relight from a lighting-sphere image, and recast people from one photo each - with limits and per-second prices
  Camera controls (R3): first frame image + camera_trajectory of keyframes {time 0-1 normalised, azimuth degrees, elevation -90..90, distance in normalised scene units}; first pose held before its time, last held to the end; signed full turns preserved up to 32 turns total azimuth; duration 0.92-15 s (output may run ~0.7 s long); 480P/768P native, 1080P is latent refinement from 768P; the default prompt freezes the scene so "only the camera moves" - i.e. a turntable/orbit tool for set or product coverage from a single still. Price (R6): promotional $0.0325/$0.052/$0.104 per s (480/768/1080p), 35% off until 31 Oct, then $0.05/$0.08/$0.16. Relight (R4): source video up to 15 s plus a sphere render showing the target lighting; video is converted to constant 24 fps and trimmed to the nearest 17n+5 frame count (minimum 56 frames, ~2.33 s), so plan for possible tail trimming in the edit; 480P/768P native, 1080P and 2K via latent upscale; $0.05-$0.32 per s. Recast (R5): source 5-30 s with no...
• articlePer YouTube Help (R1): eligibility minimums - channel in YPP, meets the YPP subscriber threshold, based in an eligible country, not a music channel / Official Artist Channel / associated with music partners (labels, distributors, publishers, VEVO), and not Made-for-Kids at channel level or with...
  Per YouTube Help (R1): eligibility minimums - channel in YPP, meets the YPP subscriber threshold, based in an eligible country, not a music channel / Official Artist Channel / associated with music partners (labels, distributors, publishers, VEVO), and not Made-for-Kids at channel level or with a significant number of MFK videos; pilots may differ; invitation is not guaranteed (interest form). Country lists on the same page disagree: the programme note lists 14 (Argentina, Brazil, India, Indonesia, Japan, Korea, Malaysia, Mexico, Philippines, Singapore, Taiwan, Thailand, US, Vietnam), while the eligibility bullet lists 18, adding Germany, Netherlands, Poland and the UK - treat the newer-looking 18-country list as unconfirmed for those four until Studio shows the programme. Join: Studio > Earn > Programs > Join Now and accept the Shopping Affiliate terms; then tag products (tagging guidelines apply). Tools: Affiliate hub (Earn > Shopping > Explore affiliate offers) shows sellers and ...
• articleWhy it matters: many earlier candidates pin prices and schemas on fal (e.g. C3-B501..C3-B507), and vendor pages there carry a "Commercial use" badge, but the governing rights are layered. From fal's terms (R1): (1) Inputs - the customer keeps ownership; fal receives a non-exclusive,...
  Why it matters: many earlier candidates pin prices and schemas on fal (e.g. C3-B501..C3-B507), and vendor pages there carry a "Commercial use" badge, but the governing rights are layered. From fal's terms (R1): (1) Inputs - the customer keeps ownership; fal receives a non-exclusive, non-sublicensable licence to process Customer Input "to provide the Services" (no training licence stated for inputs). (2) Usage Data - defined as anonymised or aggregated data "which may include data based on or derived from Customer Input", which fal owns and may use to "design, develop, and offer Company products, services, and AI models" - so the no-training reading of inputs is not absolute for derived aggregate data. (3) Outputs - no clause assigning output ownership was found; fal "does not represent, warrant, or covenant that any Output Content will be original, will not infringe rights of any third party", outputs may be the same or similar for other users, and use is at the customer's own risk;...
• articleCascadeur as an "assistive AI" 3D animation tool: local inference, documented training-data provenance, no training on user animations, video-to-pose mocap alpha, and a new MCP server hook
  Nekki's Cascadeur publishes an unusually specific AI-use disclosure (R1): ML tools are AutoPosing, Inbetweening, AutoInterpolation and Video Mocap; AutoPhysics, Ragdoll, Animation Unbaking and Fulcrum cleaning are explicitly non-ML. Pose/inbetween models were trained on Nekki's own game animation (hand-key + mocap for Shadow Fight and SPINE), supplementary mocap and purchased third-party packs "certified for AI use"; the video mocap uses open-source libraries plus post-processing. User animations are not collected for training, and all tools run locally (mocap downloads a ~700 MB mocap.pt package once). Video Mocap (Alpha, R2): import an .mp4 reference, select N timeline frames, get N keyframed poses - single actor only, poses can miss nuance because a 2D view hides depth; intended for animation drafts, then refine with AutoPosing/physics. 2026.2 (R3): Additive Layers (alpha), Easing on intervals, Inbetweening/Motion Generation retrained on larger datasets and now passing style betw...
• articleWhat measured studies say about titles and thumbnails: rewriting titles for each community lifts engagement; outrage framing lifts engagement too - at a social cost
  Two studies (abstract level). (1) Hu et al.: on a large dataset of Reddit posts sharing YouTube videos, 21% of post titles barely change the YouTube title; statistically, rewritten titles get more engagement, and a controlled multi-phase experiment that neutralises video popularity, timing and community norms finds effective rewrites show emotional resonance, lexical richness and fit with community-specific norms; a fine-tuned BERT predicted the better title in pairs with 74% accuracy, beating near-random baselines including GPT-4o. Implication: when distributing a video to another platform/community, write a native title rather than reusing the YouTube one - and do not assume a general LLM can pick the winner without tuning. (2) Park et al.: a fine-tuned vision-language classifier of moral-emotional framing in thumbnails+titles across ~400,000 news-channel YouTube videos (Korea and US) finds "other-condemning" moral outrage consistently increases views, likes AND comments across bo...
• articlePer YouTube Help (R1): the uploading channel can invite "up to 10 creators" as collaborators (some third-party coverage and a search summary on the same day cited 5, likely an earlier rollout figure - the help page text is the current reference). Works for long-form and Shorts, and for archived...
  Per YouTube Help (R1): the uploading channel can invite "up to 10 creators" as collaborators (some third-party coverage and a search summary on the same day cited 5, likely an earlier rollout figure - the help page text is the current reference). Works for long-form and Shorts, and for archived live streams, not active live streams; visible on desktop, mobile and TV; each collaborator's name and subscribe button appear only after they accept. Set it in Studio upload > Details > Audience > Show more > Collaboration > Invite a collaborator. Discovery: "Our goal is to recommend the video to the audiences of both the collaborators and the channel where the video was posted", with the caveat that changes in search or recommendations may not be noticeable at first - YouTube publishes no effect size. Revenue: the posting channel receives all earned revenue; "Revenue is not split between collaborators" - so revenue sharing must be agreed off-platform (contract or invoice), and alternating u...
• articleDid the AI-voice newscast EEG study also measure behavioural credibility, and what did it find?
  Problem. Whether synthetic voice lowers credibility of narrated news. Conditions. Frontiers in Psychology 2023 study: a pre-experiment session plus a formal EEG/ERP session (30 participants). Method. The pre-experiment was a within-subject design replicating earlier studies on news attributed to machine versus human newscasters. Result. Human voices were perceived as more credible and fluent than AI voices (credibility t(9) = -4.000, p = 0.003; comprehensibility not significant). Correction. The earlier finding should say the paper itself reports a small behavioural credibility result (df = 9, so about 10 raters; not stated separately here), consistent in direction with the EEG result, rather than implying EEG is the only credibility evidence and contrasting it with behavioural ratings in other work. Version/date. Accepted 27 Oct 2023.
• articleWhat it measures (R1): an Elo-like composite on OpenArt's own Arena, anchored at MiniMax H3 = 1,000; weights 30% general capabilities, 17.5% talking, 17.5% singing, 17.5% multi-character speaker assignment, 17.5% multilingual. Outputs are generated by OpenArt from procured prompts - "Model...
  What it measures (R1): an Elo-like composite on OpenArt's own Arena, anchored at MiniMax H3 = 1,000; weights 30% general capabilities, 17.5% talking, 17.5% singing, 17.5% multi-character speaker assignment, 17.5% multilingual. Outputs are generated by OpenArt from procured prompts - "Model providers cannot submit, select, or regenerate those outputs" - but OpenArt operates the arena and sells access to the ranked models, so the post itself says it "should not be described as an independent third party". Scores are relative within this board only; no confidence intervals are printed, so the 6-point Seedance 2.0 vs 2.0 Mini gap is not a demonstrated difference. Wan 3.0 is second on the same snapshot's Video Overall but last here - a general video ranking is a poor proxy for dialogue work. Scope gap: only five native audio+video generators are ranked; dedicated lip-replacement (sync, VEED, HeyGen, LatentSync) and image+audio avatar models (OmniHuman, InfiniTalk, Fabric) are absent, so ...
• articleHas the Commission published, in the Official Journal, any harmonised standard reference under Directive (EU) 2019/882 (EAA), and is EN 301 549 on such a list?
  Curator re-ran a read-only title query (English titles since 2021 containing '2019/882') and, like the producer, found no Official Journal act publishing harmonised-standard references under the EAA as of 2026-10-21; therefore EN 301 549 cannot be reported as cited on an EAA list.
• articleYouTube multi-language audio (creator-uploaded dub tracks) vs automatic dubbing: eligibility via Advanced features, upload steps, localized thumbnails, copyright-mismatch removal, and YouTube's ">25% of watch time" non-primary-language claim
  Per YouTube Help (R2): multi-language audio lets creators upload their own dubbed tracks (also to Shorts) and, for long-form, language-specific thumbnails; it does not generate tracks. Available to channels with Advanced features. The default track follows the viewer's preferred language (from watch history); translated titles/descriptions are used by search. Upload: Studio (desktop) > Languages > video > Add language > next to "Dub" click Add > audio-only file roughly the same length as the video > Publish. If an automatic dub exists for that language it must be deleted first. Rights warning: if the copyrighted content in a secondary track differs from the original audio, the file may be removed - so music or effects beds must match the original mix. Edits to the original (trim, audio swap) propagate to secondary tracks. Localized thumbnails: upload the main thumbnail, then add per-language thumbnails under Languages. YouTube's own claim: "Creators who uploaded Multi-language audio...
• articleWhere do the Kuleshov experiments agree and conflict on effect direction, size and rating dimension, and which reported numbers should an agent re-check before quoting?
  Agreement: negative/fear context lowers valence ratings of an unchanged neutral face in the four studies whose valence direction is stated in text (Mobbs reports a significant effect, direction in a figure). Conflict: the happy-context effect, the arousal direction (fear raises arousal in Calbi and Cao Exp 2, but unpleasant context lowered arousal in Mullennix), and the forced-label effect (present in Calbi, partial in Mullennix, absent in Szaszko & Loebus). Effect sizes range from partial eta squared about .64-.67 (Cao) to Cohen d 0.05-0.08 (Szaszko & Loebus) and are not comparable. Several published numbers do not reconcile internally (Cao Fig 3 caption, Mullennix arousal eta squared, Szaszko Table 2 'd' column), so exact figures should be taken from tables, flagged, and not mixed across papers.
• articleYouTube likeness detection (Help Center, Oct 2026): who can enrol, ID + selfie verification, face-only matching with voice rolling out, review actions (privacy removal, copyright removal, archive), parody/satire limits, and data retention
  Per YouTube Help (R1): likeness detection finds uploads where an enrolled creator's face "appears to be altered or generated by AI". Status: experimental and unavailable in some countries. Enrolment: over 18, Channel Owner or Manager; Google identity verification with a government-issued ID plus a short selfie video, which also serves as the detection reference (verification can take up to 5 days); each on-camera person enrols separately (shared team accounts cause name-mismatch failures - add individuals as Managers). Mechanism: a one-time scan of newly uploaded videos, "similar to Content ID" but for faces; other faces are detected during the scan but discarded and cannot be identified. Coverage: face only today; some users can now record a speaking-voice sample, and YouTube says it is extending detection to audio in 2026; voice misuse meanwhile goes through the privacy complaint process. Matches can include real (non-AI) reuploads of the creator's footage, which cannot be removed...
• articleIs the OpenEXR 'acesImageContainerFlag' attribute still a supported thing to write, and did OpenEXR say anything about its name or status?
  OpenEXR maintainers settled the status of the ACES container classes, not the meaning of a mixed signal. PR 2560 (merged 2026-09-05) made acesImageContainerFlag a standard attribute after finding that ImfAcesFile never wrote it. The ACES file classes and exr2aces were deprecated in 3.5.0 (2026-09-22), and the issue about the missing flag was closed on that basis. When a reviewer asked what a reader should do if the ACES flag and a different colorInteropID co-exist, no rule was given.
• articleSponsored videos: YouTube's paid-promotion checkbox adds an on-video label and unlocks age/country limits, but the US FTC says a platform tool alone may not be enough - say it in the video, repeat it in live streams, avoid "sp/spon/collab"
  YouTube (R1): any branded content, sponsorship, endorsement or other commercial relationship must be declared with the paid promotion setting in video details; this adds a disclosure label at the start of the video. Selecting it also exposes custom minimum-age and per-country age restrictions (desktop Studio, new or existing videos) for local advertising laws or partner requirements. YouTube points creators to national regulators (FTC US, ASA UK, DGCCRF France, Medienanstalten Germany, KFTC Korea) and states its article is not legal advice. FTC influencer guide (R2, Nov 2019): disclose any financial, employment, personal or family relationship and any free or discounted product even if not asked to mention it; place the disclosure with the endorsement; in video, the disclosure should be in the video itself and not only in the description; in live streams repeat it periodically; "Don't assume that a platform's disclosure tool is good enough" - use it in addition to your own disclosur...
• articleFFmpeg n9.0.2 documents bitexact as writing only platform-, build- and time-independent data, primarily for regression testing; in code it drops the library version from the encoder tag and the local time from the WAV peak chunk, which stabilises bytes but removes version provenance from the...
  FFmpeg n9.0.2 documents bitexact as writing only platform-, build- and time-independent data, primarily for regression testing; in code it drops the library version from the encoder tag and the local time from the WAV peak chunk, which stabilises bytes but removes version provenance from the file.
• articleApart from Baranowski & Hecht, what do open sound-context studies (Ansani et al. 2020; Wiking 2024) report about music or sound changing how a character or a face is judged, and how reliable are they?
  Two open sources add sound-context evidence, neither a neutral-face replication of Baranowski & Hecht. Ansani et al. 2020 (N=118 online, then N=92 in lab) put a 1 min 55 s scene from a short film with a character walking under three soundtracks (melancholic jazz piano, anxious orchestral, ambient control) and found music shifted empathy toward the character, some personality impressions, plot expectations and perceived coziness of the setting, but not the omnibus ratings of the character's affective state (only the single item 'wistful'). Wiking 2024 is a Swedish bachelor thesis with 20 volunteers and open-ended answers to silent versus sound-accompanied face clips; it reports no statistics, its methods text and analysis text disagree in places, and its main observation (child laughter read as irritation or sadness) is anecdotal. An open search found no further open-access experiment that places music under a neutral face (bounded: four OpenAlex queries, top 10 each).
• articleDo thumbnail, title and script features predict a YouTube video's views? What does measured evidence say vs creator advice?
  A Sept 2026 preprint (Tariq, arXiv 2609.32925v1; abstract read) links 3,728 thumbnails, 385 hours of speech (2.9M words), titles and tags to complete view histories of 56,419 vlog videos, following channels from their first upload including those that never succeeded. Findings as stated: 36 image/speech/text measures explain at most 1.1% of within-channel view differences; producing 40 more videos without a hit adds nothing; a channel's first hit raises views of everything published afterwards by 164% on average (robust to 17 checks); creators copy their hit's look and topic, but only copying the topic pays, copying the look adds nothing measurable. Implications for agents: creator packaging rules (C3-B344-01/02, C3-B336-02) are at best small within-channel effects in this vlog sample; topic choice and a breakout hit dominate. Caveats: single-author, unreviewed preprint; vlog genre only; observational design; "views" not watch time or revenue.
• articleCan you use Suno songs commercially? What the current Terms (effective 3 Sep 2026) actually say
  Primary terms, not legal advice. (1) Paid Pro and Premier: Suno assigns to you its right, title and interest in Output generated from your submissions, but the Output stays subject to the Terms "including any applicable commercial use restrictions", and Suno makes no warranty that copyright vests in any Output. (2) Free or Basic tier: outputs only for lawful, personal, non-commercial purposes. (3) Default rule: no commercial exploitation of the Service, Output or Voice Models except "Permitted Commercial Use": you may commercially exploit Output only if you obtained a permitted DOWNLOAD through an approved channel within your tier's monthly download allocation (shown at suno.com/account and changeable); non-downloaded Output may not be exploited commercially. Commercial rights in obtained Downloads are perpetual and survive later allotment changes. (4) Remixes made via remix features are joint works owned equally with the original creator (where copyright exists) and may only be use...
• articleDoes moral-outrage framing in titles/thumbnails raise engagement, and can misleading thumbnails be detected automatically?
  Park et al. (arXiv 2601.21815v2; abstract read) classify moral-emotion framing in thumbnail+title of ~400,000 videos from major news channels in Korea and the US and report that "other-condemning" moral outrage consistently increases views, likes and comments across both cultures - observational and limited to news outlets, with the authors warning about polarization; classifiers released. Naveed et al. (arXiv 2607.23739v1; abstract read) built a 2,843-video, eight-country dataset including 1,359 misleading-thumbnail videos (>7.6B views) and report LLM detection up to 93.8% accuracy (best: Claude 3.5 Sonnet) using thumbnail, video-to-text descriptions and subtitles. Together: outrage and misleading packaging can buy engagement, but platforms and researchers can detect misleading thumbnails, and YouTube's thumbnail policy and inauthentic-content rules apply. Agents should not recommend outrage or misleading packaging as a growth tactic.
• articlesyncnet_python (907c0b57) conforms video to 25 fps and audio to 16 kHz mono before analysis, reports an integer frame offset (vshift minus argmin of averaged distances) with confidence = median minus minimum, so pipeline offsets are in 25 fps frames (40 ms) regardless of source frame rate; code...
  syncnet_python (907c0b57) conforms video to 25 fps and audio to 16 kHz mono before analysis, reports an integer frame offset (vshift minus argmin of averaged distances) with confidence = median minus minimum, so pipeline offsets are in 25 fps frames (40 ms) regardless of source frame rate; code is MIT while the VGG page offers the model "for research purposes under Creative Commons Attribution License".
• articleWhat numeric dialogue-vs-music/background limits and monitoring methods do the BBC, Channel 4/DPP text and EBU publish for dialogue intelligibility in mixes?
  Corrected 2026-10-23 - see C3-B645-03; the original statements below are kept for history and are superseded where the correction says so. Broadcaster numbers for dialogue clarity are recommendations from different methods and are not interchangeable. The BBC (Sound Mixing guide v5.0.5) suggests pulling music down 4 dB after the final mix and monitoring on a domestic TV set to 65 dB C-weighted with 20 dB pink noise, noting that passing R128 alone does not ensure clarity. Channel 4 DPP v5.2 recommends at least 4 LU dialogue-to-background separation and a factual dialogue LRA of 6 LU or less, as guidance only. EBU R 128 s4 (2023) says the loudness-to-dialogue ratio of cinematic content should not exceed 5 LU (15 LU or more occurs). A Dutch process cited in an EBU review caps voice-to-programme loudness at 4.5 LU.
• articleHow does YouTube count unique viewers, and why is subscriber count a poor audience-size measure?
  YouTube Help (captured 2026-10-23): unique viewers is an estimate of distinct people who watched in a period; one person on several devices or watching repeatedly counts once, and the model accounts for shared devices; totals are available for windows up to 90 days, from 2017-08-01; for Shorts, unique viewers are estimated from engaged views (people who stayed to watch). A viewer of several videos counts once per video but once at channel level. YouTube says subscriber counts are not an accurate audience-size measure because viewers subscribe to dozens of channels and do not return for every upload. Completes the missing part of CA-DR-0001-03 (C3-B336-03).
• articleWhat do creator tests say Kling 4.0 does at launch, and which headline claims are not yet available?
  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. Two Mikkel Lassalle videos (2026-10-20 promotional showcase linking kling.ai; 2026-10-22 Artlist-affiliate test). Claims: generations up to 30 s at up to 1080p; multiple keyframes (Artlist video: up to 10) to fix states along a long shot; omni/multi-reference (Artlist: up to 15 image+video references, matching fal schema 10 images + 5 videos in C3-B388-01); stereo audio with lip-sync and dialogue in nine languages; much better in-video text, emoji and logo rendering; 21:9 output; Kling 4.0 Flash up to 20 s at 720p. The first video's TITLE advertises "10-Bit HDR, 4K" but its narration says Kling 4.0 "will also be able to support 10 bit HDR, 4K ... and video extension" on...
• articleHow complete is the Croissant that Hugging Face auto-generates for licence and provenance?
  Corrected 2026-10-23 - see C3-B644-04; the original statements below are kept for history and are superseded where the correction says so. For the four datasets checked (a sample, not a rate), Hugging Face's auto-generated Croissant carries a single dataset-level licence URL copied from the card's licence tag (choosealicense.com form), with no citation, provenance or responsible-AI fields filled, no per-file licences, and a GitHub-issue URL in place of each file's sha256. A dataset without a card licence tag (common_voice_17_0, now distributed via Mozilla Data Collective) gets plain schema.org JSON with no licence or distribution. Auto Croissant is therefore not a provenance or rights record.
• articleElevenLabs Music: what each plan lets you do with generated songs - no film, TV, radio or studio games below Enterprise Music (terms of 9 Oct 2026)
  Primary terms (Exhibit A). Media Rights for Free, Starter, Creator, Pro, Scale, Business and Enterprise Music Lite: "All online and offline commercial use permitted, except film, TV, radio, and Studio Games"; only Enterprise Music (contact sales) permits all commercial use. So a filmmaker scoring a film or TV piece with Eleven Music needs Enterprise Music. Studio Game vs Indie Game: indie = under $500,000 lifetime game revenue AND the developer/publisher under $1,000,000 annual game revenue (net, incl. affiliates); if a game stops qualifying you have 60 days to upgrade or stop using the output (already-distributed copies may continue). Other table rows: eligibility - Free to Pro individual use only, Scale < 10 employees, Business < 50; monthly generation 11 / 17 / 62 / 304 / 1,100 / 4,800 minutes (Free..Business); monthly downloads not permitted on Free, then 30 / 250 / 500 / 1,500 / 4,000 minutes; streaming distribution prohibited on Free and Starter, permitted from Creator up; mus...
• articleWhich AI music models produce the most preferred songs with vocals? Artificial Analysis vocal board snapshot (23 Oct 2026)
  Closes the open follow-up on the vocal music board (the instrumental board was captured earlier). AA-Music-Vocal v1.1, pairwise human-preference Elo anchored at MiniMax Music 3.0 = 1000, about 2,600-2,800 samples per model, 95% intervals about +/-19-20. Ranking: 1-2 Suno v6 1147 and Suno v6-mini 1130 (statistically tied top pair); 3 Mureka V9.5 1106; 4-6 Mureka V9 1059, Lyria 3 Pro 1051, Lyria 3.5 1051; 7 Eleven Music v2.5 1013; then a tied cluster MiniMax Music 3.0 1000, Eleven Music v2 997, MiniMax Music 2.6 997, Eleven Music 997, MiniMax Music 2.5+ 987; Melodia v3 970. Practical reading: rank ranges matter more than single ranks - the cheaper Suno v6-mini is statistically indistinguishable from Suno v6 here; Lyria 3.5 did not beat Lyria 3 Pro. Preference is not licensing: commercial-use terms and training-data/copyright risk differ by provider (see creator observation of memorised-sounding songs in C3-B393-02) and must be checked separately; prices not captured on this page.
• articleKling 3.0 native 4K and multi-shot: claimed uses (pickup shots, documentary inserts) and how to read "world first" claims
  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. Mikkel Lassalle (Kling promotional, 2026-05-07). Kling 3.0 with native 4K output: one prompt with the multi-shot toggle yields several angles with consistent characters; workflow = meta-prompt (ask ChatGPT for a detailed documentary-style prompt with a reference prompt) -> generate in 4K; for more control generate an image first, then image-to-video; create extra angles of the same scene via image-to-image ("drone shot of the lion") and animate them. Suggested production use: generate a missed close-up instead of a reshoot. Claims "World's First Native 4K AI Model" (title) are marketing; not verified. Cross-check: Runway lists kling3.0_4k at 43 credits/s ($0.43/s) and f...
• articleHow does Europeana's own rights-statement set relate to RightsStatements.org, and what should an agent expect to see in Europeana records?
  Europeana accepts 14 rights values: Creative Commons licences and tools plus six RightsStatements.org statements (InC, InC-EDU, InC-OW-EU, NoC-NC, NoC-OKLR, CNE); InC-RUU, NoC-US, NoC-CR, UND and NKC are not in its list. It states rights statements are not licences, discourages CNE (the RS vocabulary itself describes CNE neutrally) and requires providers' best efforts, warning that a wrong public-domain label misleads users. RS JSON-LD closeMatch links point at legacy Europeana rights URIs (InC to europeana.eu/rights/rr-f/, CNE to /rights/unknown/, InC-OW-EU to /rights/orphan-work-eu/; InC-NC to CC BY-NC 4.0), so legacy and RS URIs may both appear in older data.
• articleHow does an agent read a RightsStatements.org URI as data, and which machine-readable properties does each statement actually carry?
  Each of the 12 RightsStatements.org 1.0 URIs returns JSON-LD under content negotiation (all dated 2022-04-27, version 1.0). The machine layer is thin: closeMatch links on 4 statements, relatedMatch on 10, and a CommercialUse prohibition only on InC-NC and NoC-NC; InC-EDU's educational permission exists only in prose (an ODRL prefix is declared but unused). The technical white paper says URIs are opaque to machines and requires HTML5+RDFa, JSON-LD and Turtle representations. So software can classify a status but cannot read most permissions from the data.
• articleWhat do the LGPL 2.1 and GPL 2 texts and FFmpeg's checklist say about linking and distributing FFmpeg in a media pipeline, and what do they leave unaddressed?
  The licence texts shipped with FFmpeg say: under LGPL 2.1 a program that links the library is a 'work that uses the Library' and may be distributed on the distributor's own terms if they allow customer modification and reverse engineering for debugging and provide either relinkable materials/source or use a shared-library mechanism; under GPL v2 a work containing GPL code must be licensed as a whole with corresponding source. Both say running the program is not restricted and its output is covered only if it is itself a work based on the program. FFmpeg's checklist offers one conservative route (no --enable-gpl or --enable-nonfree, dynamic linking). The texts do not discuss subprocess pipelines or hosted services, and the project declines to answer patent questions.
• articleIs the OpenEXR prose-versus-flowchart discrepancy (attribute 'present' versus flag 'set') reported in a public ColorInterop tracker issue, and what do the issues say about it?
  The ColorInterop OpenEXR recommendation is internally inconsistent on ACES container detection, and this was still true on main on 2026-10-22. The reading steps say the acesImageContainer attribute takes precedence if it is present; the flowchart asks whether the flag is set. The name also differs from OpenEXR's: OpenEXR's PR 2560 (merged 2026-09-05) added the standard attribute as 'acesImageContainerFlag', not 'acesImageContainer'. Readers should test for the OpenEXR attribute name and decide explicitly whether presence or a true value counts.
• articleWhat are IMF, DCP, ProRes and FFV1/Matroska each for (distribution master, cinema, editing mezzanine, preservation)?
  Format roles differ by purpose: FFV1 (RFC 9043) is lossless intra-frame coding suited to preservation or intermediate use; the Library of Congress prefers, for finished films, an unencrypted DCP (24/48p, at least 2048x1080) among other forms and ProRes 4444(XQ)/4444/422 HQ, and for file-based video lists IMF packages, FFV1 v3 in Matroska and ProRes in that order. Ask whether the file is for hand-off, screening, editing or archive before choosing.
• articleWhat does Netflix Photon actually validate for IMF (ST 2067 / App #2E), and what is left unchecked?
  Netflix Photon parses and validates IMF packaging (AssetMap, PackingList, CPL, track files) and application constraints for App #2E and App #5 ACES; its App #2E validator checks that images are JPEG 2000 within IMF 2K/4K/Broadcast profile resolutions, HTJ2K sub-descriptor and RPCL progression, and permitted colorimetry/sampling/quantisation combinations. In the files read there are no loudness or picture-quality checks, so Photon conformance is structural, not a QC pass.
• articleHow do ITU-T P.800 and P.808 define the MOS scale, listener rules and crowdsourcing controls, and how do the two Recommendations differ?
  ITU-T P.800 defines the listening-only ACR test whose 5-point mean is the MOS (with a note that older results must be shifted by one to compare) and requires naive listeners (no subjective test in 6 months, no listening-opinion test in a year); P.808 (in force, 2021) adapts ACR/DCR/CCR to crowdsourcing with one gold-standard and one trapping question per 10 stimuli, two-eared headphones and a two-week listening-test gap, and says crowdsourcing complements rather than replaces laboratory tests. Neither defines a TTS-naturalness protocol.

More