Evaluation and media provenance

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

• articleHugging Face model cards: lineage, licence and version metadata
  Check base_model (and base_model_relation) to tell an original model from a fine-tune, adapter, quantization or merge; license: other means read license_name and license_link; new_version links to a newer model.
• articleHow easily can invisible image watermarks be removed, and which watermark types resist regeneration or adversarial attacks?
  Corrected 2026-10-24 - see C3-B1132-01; the original statements below are kept for history and are superseded where the correction says so. Corrected 2026-10-23 - see C3-B681-04; the original statements below are kept for history and are superseded where the correction says so. Two benchmark-style papers test removal. WAVES (5 watermarks, 26 attacks) finds no attack works on all watermarks: regeneration/diffusion attacks are most effective on Stable Signature (average detection performance 0.001 under Regen-Diff), Tree-Ring is most vulnerable to adversarial attacks, and StegaStamp is most robust but visibly degrades images. Zhao et al. report over 99 percent removal for DwtDctSvd, RivaGAN and SSL by regeneration, with StegaStamp needing heavy noise. The two papers disagree on Tree-Ring under regeneration, probably because of different strengths and metrics, and both use attackers without detector access.
• articleWhy does an autoregressive video model get worse the longer it keeps generating, and what are the names and visible symptoms of the failure?
  Corrected 2026-10-24 - see C3-B1094-02; the original statements below are kept for history and are superseded where the correction says so. The primary papers agree on one mechanism: the model is trained on clean or short context but, when it keeps generating, it conditions on its own slightly wrong frames, so small errors compound (called error accumulation or exposure bias). Reported symptoms differ by method: brightness drift (over-exposure or darkening), motion that freezes, colour shift, flicker when a window shifts, and (separately) content that diverges because nothing remembers what left the frame.
• articleWhat evaluation protocol (resolution, clips, frames, FID/FVD scope, quoted versus recomputed baselines) does each later paper state for its nuScenes FID/FVD, so which tables can even be set side by side?
  Corrected 2026-10-24 - see C3-B1051-01; the original statements below are kept for history and are superseded where the correction says so. The stated protocols differ in at least five ways: Drive-WM 150 validation clips x 6 views = 900 segments of 40 frames with the official UCF FVD code; GenAD FID on 6019 frames and FVD on 4369 clips; UniMLVG following MagicDrive with 900 videos of 16 frames; Panacea generating the whole validation set at 256x512 with 8 frames; Delphi single-frame single-view FID and 4/8/40-frame FVD with recomputed baselines marked. DriveDreamer v1 states a 192x128 training size, v2 448x256 for training and evaluation; MagicDrive-V2 lists DriveDreamer at 128x192 with 32 frames. DriveDreamer-2 gives no clip or frame count for FID/FVD in the HTML text searched, and DriveDreamer v2 has no clips/segments sentence at all; DriveDreamer v1 states 150 clips x 8 segments = 1200, but for action-prediction evaluation. DrivingDiffusion says it matched metrics and settings "a...
• articleWhat do reading experiments show happens to a character's goal in the reader's mind when the text repeats it, replaces it with another goal, or says nothing?
  Corrected 2026-10-24 - see C3-B1074-01; the original statements below are kept for history and are superseded where the correction says so. In one open-access experiment (Linderholm, Gernsbacher, van den Broek et al., 2004; 84 retained undergraduates, 32 three-paragraph texts) readers verified a goal word faster after the goal was restated (713 ms) than after a neutral paragraph (834 ms), and slower after a new goal replaced it (908 ms), with matching error rates. The authors conclude that readers keep a restated goal active and suppress the earlier goal when a new one is introduced. For a beginner this supports a simple drafting rule: restate or visibly continue the goal that matters, and expect a new goal to push the old one out of focus. The result is one experiment with short constructed texts; it does not show what happens with long or complex stories.
• articleWhich terms govern the text-safety models bundled with or called by Cosmos-Transfer1: meta-llama LlamaGuard-7b, NVIDIA Aegis Defensive 1.0 and Llama-Guard-3-8B?
  Corrected 2026-10-24 - see C3-B1125-02; the original statements below are kept for history and are superseded where the correction says so. The HF Cosmos-Transfer1-7B snapshot at ef4dc8e3 caches meta-llama/LlamaGuard-7b (Llama 2 Community License per Meta's gating text) and two revisions of NVIDIA's Aegis adapter (card: governed by the Llama 2 Community License). The GitHub code at 5005e823 instead calls meta-llama/Llama-Guard-3-8B (Meta gating text: Llama 3.1 Community License; NVIDIA ATTRIBUTIONS.md embeds the Llama 3 text), which is not in the bundle. The upstream Meta licence and use-policy files were not copied into the bundle snapshot, and the Meta READMEs and licence files are gated; the licence text comes from the public gating-prompt metadata. No legal conclusion is drawn.
• articlePROPOSAL (1916 practice as the basis): what checks could an agent run on a draft scene list for plot-by-action, continuity and establishment, using only what Sargent's Chapters XLIII-XLV state, and what in the check is only the lead's or worker's addition?
  Corrected 2026-10-24 - see C3-B1088-01; the original statements below are kept for history and are superseded where the correction says so. Agent check, steps 1-8 labelled '1916 practice' (each tied to a Sargent claim in this candidate's evidence) and steps 9-10 labelled PROPOSAL (worker additions, not in Sargent). The check reads a numbered scene list, locates the first vital action, the premise treatment, the establishing scene for each plot role, and each scene break's time, room, dress and entrance. It reports violations by rule and scene number. Sargent gives no test, no measured threshold and no props rule; the only criterion he states for casual setup is the 'reason-back' contrast. The check has not been run. It is not a validated detector and the rules are 1916 silent-film conventions.
• articleI have a draft with an intended theme: what can I check to see whether a reader could recover it, and what do the sources say about how far such a check can be trusted?
  Corrected 2026-10-24 - see C3-B1069-01; the original statements below are kept for history and are superseded where the correction says so. No inspected source validates a theme-recoverability test for drafts; what exists are building blocks. Theme in the Seifert/Dyer line is a goal-plan-outcome pattern; people sorted stories by such patterns in studies reported second-hand; Levine's literal-vs-interpretive rubric reached 84% agreement between two raters; and expert readers give related but different theme statements. The check below (state theme as pattern, blind restatement by two or more readers, code with the rubric, report agreement, allow variants) is this worker's synthesis, not a measured method.
• articleIf the user lives in the EEA, Switzerland or UK, which ElevenLabs Terms of Service apply, and do the voice-model and consent clauses differ from the non-EEA Terms?
  Corrected 2026-10-24 - see C3-B1128-01; the original statements below are kept for history and are superseded where the correction says so. Corrected 2026-10-23 - see C3-B678-04; the original statements below are kept for history and are superseded where the correction says so. ElevenLabs publishes two Terms of Service, both 'Last Updated: 09 October 2026': one for residents of the EEA, Switzerland or UK and one for everyone else. Section 4 (Content and User Voice Models) has 27 text blocks (heading and paragraphs) in each; after removing bullet glyphs 26 are word-for-word equal, including the 'voice you are authorized to share with us' wording, the content/voice-model licences and the training opt-out. The one differing paragraph is 4(h): non-EEA bars HIPAA protected health information, the EEA version bars Article 9 GDPR special-category data including biometric data for unique identification. The rest of the differences (withdrawal right, consumer liability, no arbitration, consu...
• articleWhat counts as a plot hole or continuity error in a story, and how do open papers define and categorise them so I can tell which kind I am looking for?
  Corrected 2026-10-24 - see C3-B1063-01; the original statements below are kept for history and are superseded where the correction says so. Two open papers give working definitions an agent can apply. FlawedFictions (v3) defines a continuity error as a statement contradicted by what else the story establishes, within five plot-hole kinds it takes from a writing source; ConStory-Bench defines consistency errors as contradictions between a fact stated earlier and text later, in 5 categories and 19 subtypes. Both treat intended ambiguity and genre conventions as outside the definition, and neither measures craft judgement about whether a hole matters.
• articleWhat does MPJPE measure, and which public monocular pose benchmarks actually use marker-based optical ground truth?
  Corrected 2026-10-24 - see C3-B1090-01; the original statements below are kept for history and are superseded where the correction says so. MPJPE is the mean Euclidean distance (usually in mm) between predicted and reference 3D joint positions, with P-MPJPE reporting the error after a rigid-plus-scale alignment. Of the common benchmarks, only Human3.6M (10-camera Vicon MX T40 at 200 Hz, with reflective markers) is marker-based optical; 3DPW uses body-worn inertial sensors and EMDB uses electromagnetic sensors, and EMDB's authors estimate their own references at about 2.3-2.6 cm PA-MPJPE. So "MPJPE against marker-based ground truth" is a Human3.6M-style statement and is not interchangeable with in-the-wild numbers.
• articleHow reliably do human annotators locate acts or turning points in films and screenplays, and do their positions match the 25/50/75 percent rules in structure books?
  Corrected 2026-10-24 - see C3-B688-03; the original statements below are kept for history and are superseded where the correction says so. Viewers segmenting seven films into large parts (20-35 min units) agreed with each other and with critics' segmentations above chance but far from uniformly (Cutting and Armstrong 2019, CC BY). In TRIPOD's 99 screenplays, annotated turning-point means (11.4, 31.9, 50.7, 74.2, 89.4 percent) sit near but not on textbook positions (10, 25, 50, 75, 94.5) with standard deviations of 4.7-12.2 points, and two annotators agreed exactly on only 35.48% (partial 56.67%) at screenplay level.
• articleHiggsfield (consumer/creator tiers): you keep Outputs and may use them commercially, but inputs and outputs may train Higgsfield models, outputs may not train any other model, and the underlying model can be swapped without notice
  Corrected 2026-10-24 - see C3-B1096-01; the original statements below are kept for history and are superseded where the correction says so. Primary Higgsfield Terms of Use (last updated 26 Jul 2026). Ownership: Company "does not claim ownership" of Inputs or Outputs and does not restrict commercial use; output rights survive cancellation or account deletion and can be sublicensed to clients (s4.4). Training: Your Content, Inputs and Outputs "may be used by Company to train" its and affiliates' models; opt-out is only by deleting content or the account, going forward; Enterprise Agreement customers are excluded from training and treated as confidential (s4.4). Marketing reuse only for content you made public on the service or with consent. Likeness: uploading a person's name/likeness/voice means you represent you hold all releases and consents, including for generating Outputs (s4.2(b)). Outbound restriction: you may not use the Service, Outputs or derived data to train, fine-tune or...
• articleIn Kang et al.'s controlled 2D testbed, how do in-distribution, out-of-distribution and combinatorial generalisation differ for video diffusion models, and what can not be concluded about commercial video models?
  Corrected 2026-10-24 - see C3-B680-01; the original statements below are kept for history and are superseded where the correction says so. Kang et al. (ICML 2025) train diffusion-transformer video models from scratch on Box2D/PHYRE simulations so the training set is fully controlled. Within the training range the models track the simulator almost exactly; outside it (new speeds or radii) error is about 35x higher and does not shrink with 100x more data or 14x more parameters. Where every object type is seen but not every combination, scaling the number of combination templates (6 to 60) cut human-labelled 'abnormal' out-of-template videos from 67% to 10%. Attribute-pair tests show the models copy the nearest training example, preferring colour over size over velocity over shape. The setting is a 2D toy world, so the paper does not measure Sora-class or other commercial models.
• articleWhat does YouTube's own text say changes in the YouTube Partner Program from 1 Feb 2027 (new-creator thresholds, who is affected, terms deadline), and what is still not stated?
  Corrected 2026-10-24 - see C3-B968-05; the original statements below are kept for history and are superseded where the correction says so. Problem: C3-B566-01 recorded only the help-page banner. Primary text now read: the YouTube Official Blog post 'New opportunities to earn and changes to the YouTube Partner Program' (The YouTube Team, 10 Aug 2026), plus the YPP overview Help page. Source statements: (1) New creators applying for YPP will need 8,000 qualified watch hours in the last 365 days, or 20 million qualified Shorts views in the last 90 days, for ads and Premium revenue sharing. (2) 'This update won't impact creators already in YPP.' (3) Entry thresholds for Fan Funding and shopping products remain unchanged. (4) Creators can review and sign the new terms in YouTube Studio; the terms take effect 1 Feb 2027. (5) The Help page banner says to review and accept the updated terms in Studio by 31 Jan 2027 to continue fully monetizing. Not stated or not verified (checked 2026-10-23...
• articleWhat inputs and limits does YouTube document for AI Remix (reimagine a Short with prompts and photos), and which rules still apply?
  Corrected 2026-10-24 - see C3-B970-03; the original statements below are kept for history and are superseded where the correction says so. Problem. AI remix changes someone else's Short with generated video. Conditions. Eligible Shorts only; English prompts only; rollout described as 'coming weeks' on an undated page. Method. In the app tap Remix on an eligible Short, enter a custom prompt or pick a suggestion, optionally upload up to 3 reference photos, wait for generation and preview; audio is generated from the prompt and scene; further edits with Shorts tools before upload. Tip from the page: include a subject and at least one reference photo, and be descriptive if specific audio is wanted. Community Guidelines apply to generated output. Failure modes. Vague prompts. Limitations. No length, watermark or opt-out detail on this page (see C3-B694-10 for guardrails). Version and date. Undated; fetched page.
• articleDoes CoSER (arXiv 2502.09082) include a human evaluation of LLM-generated character dialogue, and was its LLM judge checked against humans?
  Corrected 2026-10-24 - see C3-B679-01; the original statements below are kept for history and are superseded where the correction says so. Provisional lead, not verified from raw bytes. The pinned abstract page (v3, observed 2026-10-22) shows CoSER trains open models that sequentially portray several book characters in a scene and proposes its own evaluation protocol; the abstract does not mention human raters. A condensed read of the full HTML v3 (not quotable, no raw copy because of the batch storage stop) reports a human study in Section 5.3 with a small panel scoring LLM-simulated scenes on an overall 1-10 scale and a table comparing automatic judges with the human results. The rating criterion was not identified as character fidelity, distinguishability or voice, and no agreement statistic was seen. If confirmed from raw bytes, S52's negative narrows further: humans did rate LLM-generated character dialogue, but not as who-said-it attribution.
• articleWhat do ComfyUI-Manager's security_level, install flags and model-download gate actually permit, and where do docs and README disagree?
  Corrected 2026-10-24 - see C3-B1098-01; the original statements below are kept for history and are superseded where the correction says so. Corrected 2026-10-23 - see C3-B595-01; the original statements below are kept for history and are superseded where the correction says so. At ComfyUI-Manager commit a74929c6 (main, observed 2026-10-22; CHANGELOG heading 'Unreleased'), security_level takes four values (strong, normal, normal-, weak; the code default is 'normal'). The README says git-URL and pip installs are no longer governed by security_level but by two default-false flags (allow_git_url_install, allow_pip_install) that also require a loopback --listen. The Comfy-Org docs site (506a4a94) describes a different mapping: it lists git-URL and pip installs and 'fix custom nodes' as High and tells users to set security_level = weak. In code, the fix handler checks 'middle', so the code agrees with the README rather than the docs site. Model downloads through install_model pass three g...
• articleLuma's public API in Oct 2026 still documents only Ray 2 / Ray 2 Flash - while ComfyUI deprecated its Ray 2 nodes and Luma is absent from current AA video boards
  Corrected 2026-10-24 - see C3-B976-01; the original statements below are kept for history and are superseded where the correction says so. Primary API docs list two video models - Ray 2 (ray-2) and Ray 2 Flash (ray-flash-2) - with text-to-video and image-to-video via start/end keyframes, resolutions 540p/720p/1080p/4K, durations such as 5 s, loop and aspect-ratio options, and "concepts" (camera-motion presets) retrievable from /generations/concepts/list; the FAQ says API outputs carry no watermark and may be used commercially. Signals that this surface is ageing: ComfyUI deprecated its Luma Ray 2 partner nodes in v0.39.0 (5 Oct 2026, R3), and Luma models do not appear in the Artificial Analysis T2V/I2V boards captured today (C3-B438-01). We did not find (in this pass) API documentation for a newer Luma video model; Luma's consumer app may differ. Practical: before building on Luma's API, confirm model roadmap and support; older archive guidance on Luma camera concepts/keyframes (Ray...
• articleWhat do the vendors' own docs promise about seed reproducibility, and which conditions (watermark setting, session, mode, order) break it?
  Corrected 2026-10-24 - see C3-B976-02; the original statements below are kept for history and are superseded where the correction says so. The five vendors state five different strengths of seed promise. Imagen on Vertex AI is the strongest but conditional: same seed gives the same images, only when addWatermark is false, and image order is not guaranteed. Adobe's Firefly API says the same seed, prompt and presets give the same image. Midjourney calls seeds the initial noise, says they are 99% identical in V8.X, not reliable across sessions and not for Turbo mode. Ideogram says only that seed is for reproducible generation and returns the seed used. The OpenAI image guide has no seed parameter in the text read.
• articleOpen video-to-audio (AI foley) models differ mainly by licence: MMAudio weights are non-commercial, HunyuanVideo-Foley excludes the EU/UK/South Korea, ThinkSound is listed Apache-2.0 - check before adding sound to client work
  Corrected 2026-10-24 - see C3-B1037-05; the original statements below are kept for history and are superseded where the correction says so. For adding synchronized sound effects to silent generated clips (many video models still output silent video or weak foley), three widely used open V2A models carry very different rights: MMAudio - code MIT, but "the checkpoints are licensed under CC BY-NC 4.0 and can be used for non-commercial purposes only" (R1); HunyuanVideo-Foley - Tencent Hunyuan Community License, which "DOES NOT APPLY IN THE EUROPEAN UNION, UNITED KINGDOM AND SOUTH KOREA" (R2; same family of terms as other Hunyuan releases, C3-B460-01); ThinkSound - Hugging Face metadata lists apache-2.0 (R3; GitHub licence field unset, so confirm the LICENSE file). Research direction (R4-R6): WorldSonus adds real-time (real-time factor 0.41), prompt-steerable, stereo spatial audio to interactive world models via streaming causal diffusion - relevant to the silent outputs of world models ...
• articleBefore loading a shared ComfyUI workflow, what should an agent check, and which checks are documented versus only proposed?
  Corrected 2026-10-24 - see C3-B1070-01; the original statements below are kept for history and are superseded where the correction says so. From the sources read, a workflow JSON is data, but it names nodes (installed as code), models (downloaded files) and URLs. Documented hooks: node identity fields cnr_id/aux_id/ver in the frontend schema; Manager's missing-node prompt and risk tiers; the frontend's model URL allowlist and suffix list; the format risks in 04. Hash fields exist but are optional. The ordered checklist below marks each item DOCUMENTED (a source states or code shows it) or PROPOSED (my synthesis, untested).
• articleWhen must a YouTube creator disclose AI-generated or altered content, and how is it labelled?
  Corrected 2026-10-24 - see C3-B1072-01; the original statements below are kept for history and are superseded where the correction says so. YouTube requires disclosure (Studio upload > Attributes > "AI use") for realistic AI-generated or meaningfully altered content: making a real person appear to say or do something, altering footage of a real event or place, generating a realistic scene that did not occur, or creating music that is the main focus. Not required: clearly unrealistic content, minor/aesthetic edits, and production assistance (outlines, scripts, thumbnails, titles, captions, upscaling, cloning one's own voice). Labels appear in the description or on the player (music and photorealistic content). YouTube may auto-label content made with its GenAI tools, content carrying C2PA metadata, or content its systems detect; creators can usually correct it. The page states disclosure alone does not limit reach or monetization. Implication for AI music videos (frequent in creator ...
• articleCan an agent run these action-following benchmarks itself: what code, data and checkpoints does each need, in which file formats, and what do the repositories say about terms?
  Corrected 2026-10-24 - see C3-B1046-02; the original statements below are kept for history and are superseded where the correction says so. Per component at the pinned commits: GameWorldScore (code Apache-2.0 text in its directory, MIT at root; needs five external helper checkpoints in .pth/.model/.weights formats); WorldScore (code MIT, dataset on HF with metadata mit and a WorldScore-Dataset.zip in its file list, evaluation checkpoints .pth/.pt/.pkl fetched by wget, leaderboard submission by results or by videos); MineWorld (code MIT, metrics/ folder, .ckpt checkpoints named in the README tree but taken down per the README and the HF API returned 401 on 2026-10-22); CameraCtrl (Apache-2.0 code, .ckpt weights on HF); Yume (Apache-2.0 code, weights on HF per README, processed dataset listed as not yet provided, Yume-Bench data not located); GameCraft (Hunyuan Community License excluding EU, UK, South Korea; .pt weights; no evaluation code in the root listing). No repository text rea...
• articleWhat are Kling 4.0's documented limits (duration, keyframes, references, resolution, HDR, Flash), and do they match creator videos?
  Corrected 2026-10-24 - see C3-B1032-02; the original statements below are kept for history and are superseded where the correction says so. Kling's own site (kling.ai, rendered 2026-10-23) announces "All-New Kling 4.0" but the release-notes URL redirected to the homepage and no spec table was retrievable: official spec not pinned. fal.ai (a host relaying Kling's spec sheet) lists: 3-30 s per generation; up to 10 keyframe images; Omni Reference up to 15 total references (images, videos, voice references, elements), reference videos up to 5 totalling 30 s; edits with up to 5 input clips; 720p/1080p/4K output with stereo audio; 10-bit HDR at 1080p/4K "coming soon"; languages incl. Chinese, English, Japanese, Korean, Spanish, Portuguese, German, French; Kling 4.0 Flash in limited early access from 2026-09-28 (20 s/720p per other secondary pages, not confirmed here). This matches the creator claims in C3-B334-02 (10 keyframes, 15 references, 30 s, HDR pending) except his "nine languages"...
• articleOpen world models for worldbuilding (Oct 2026): persistent 3D worlds (HY-World 2.0) vs memory-based video worlds (WorldCrafter) vs keyboard-controlled video (H3-World)
  Corrected 2026-10-24 - see C3-B1036-01; the original statements below are kept for history and are superseded where the correction says so. Three open releases with different output types. (1) HY-World 2.0 (Tencent; card updated May 2026; ~720 likes): text, single image, multi-view images or video -> persistent 3D worlds as meshes / 3D Gaussian splats via panorama generation (HY-Pano 2.0) -> trajectory planning (WorldNav) -> world expansion (WorldStereo 2.0) -> composition (WorldMirror 2.0 + 3DGS); WorldMirror 2.0 (~1.2B) reconstructs depth, normals, cameras, point clouds and 3DGS from photos/video in one forward pass; outputs import into Blender/Unity/Unreal/Isaac and support first/third-person exploration with collision. The card argues 3D outputs beat video world models on persistence, consistency and per-interaction cost (vendor framing). LICENCE: Tencent HY-World 2.0 Community License does not apply in the EU, UK or South Korea, forbids using the works OR THEIR OUTPUT outside t...
• articleEU AI Act Article 50 for creators (in force 2 Aug 2026): who must label deepfakes, the lighter regime for evidently artistic or fictional work, and why "personal use" rarely covers monetised channels
  Corrected 2026-10-24 - see C3-B1049-01; the original statements below are kept for history and are superseded where the correction says so. Primary (Commission guidelines, 20 July 2026, non-binding; CJEU has final say). Four duties: 50(1) providers tell people they interact with AI; 50(2) providers mark synthetic image/video/audio/text outputs in a machine-readable, detectable way (exception for assistive standard editing that does not substantially alter input); 50(3) emotion recognition/biometric categorisation notices; 50(4) DEPLOYERS must disclose deep fakes (and AI text published to inform the public on matters of public interest, unless under human editorial control). Applies from 2 August 2026; disclosure must be clear and distinguishable at the latest at first exposure and meet accessibility requirements (50(5)). Personal-use carve-out (Art. 2(10)): only "purely personal, non-professional" use by natural persons is excluded; any activity yielding economic benefit on a regula...
• articleWhat consistency and quality evaluations have the authors of HunyuanWorld 1.0, WonderWorld, WonderJourney and World Labs published, and can their numbers be compared?
  Corrected 2026-10-24 - see C3-B1039-01; the original statements below are kept for history and are superseded where the correction says so. Each source evaluates differently and none measures 3D geometric consistency directly. HunyuanWorld 1.0 reports no-reference image-quality metrics (BRISQUE, NIQE, Q-Align) and CLIP alignment on rendered views against baselines, with no ground truth; WonderWorld's only quantitative table is generation time (9.5 s vs 749.5 s and 798.1 s on an A6000) plus qualitative comparison and a stated failure mode; WonderJourney reports a human preference study (400 participants, binary choices on four axes). The World Labs pages read give examples and claims but no published metric; its TOS disclaims dimensional accuracy. The tables are not comparable across papers (different views, baselines, protocols).
• articleWhat do the BBC File v6.0.2 (2026) and Channel 4/UK DPP File v5.2 delivery specifications say about up-converted, lower-resolution and frame-rate-converted material?
  Corrected 2026-10-24 - see C3-B1042-03; the original statements below are kept for history and are superseded where the correction says so. Both specs accept up-converted SD archive provided it meets the full technical requirements; both cap lower-resolution material at 25 percent of programme duration unless agreed; both require motion-compensated standards conversion and forbid simple timeline conversion. The Channel 4 spec adds that up-converted SD must look no worse than the original, requires removing half-line and blanking artefacts (hence a slight zoom), and requires agreement in advance for 50p to 25p film-motion conversion. In the passages read neither spec has a field or rule for declaring up-conversion; neither mentions AI (bounded: BBC text searched for "declar", "artificial", "generative" and the "AI" token with 0 hits; Channel 4 text searched for "declar" with 0 hits).
• articleWhich open image-to-3D models can studios use commercially, and are there territorial limits (TRELLIS.2, TripoSR, Hunyuan3D)?
  Corrected 2026-10-24 - see C3-B1038-04; the original statements below are kept for history and are superseded where the correction says so. Checked 2026-10-23: Microsoft TRELLIS.2 (GitHub MIT, pushed 2026-07-10; HF TRELLIS.2-4B card mit) and VAST TripoSR (MIT) are permissive. Tencent Hunyuan3D 2.1 (and 2.0, card "tencent-hunyuan-community") community licence states in capitals that it "DOES NOT APPLY IN THE EUROPEAN UNION, UNITED KINGDOM AND SOUTH KOREA"; the licence grant is "for the Territory only"; use, reproduction, distribution or display of the Works, Output or results outside the Territory is unlicensed; and licensees whose products exceeded 1 million monthly active users must request a licence. The same pattern appears in MiniMax H3 (also excluding the US) (C3-B364-01). ComfyUI ships official tutorials for TRELLIS.2, Hunyuan3D-2/2.1, Pixal3D and TripoSplat (docs index captured in C3-B370), so availability in a tool is not a licence grant. For hosted generators (Meshy, Rodin)...
• articleWhere does YouTube show its AI label since May 2026, which content gets the prominent label, and does a label change recommendations or monetization?
  Corrected 2026-10-24 - see C3-B968-03; the original statements below are kept for history and are superseded where the correction says so. Problem: creators fear that an AI label hurts reach or earnings. Conditions: YouTube, photorealistic or meaningfully AI-altered content versus unrealistic or lightly edited content. Method (platform statements): from May 2026 a single prominent label format applies to photorealistic and meaningfully AI-altered or generated content: below the player and above the description for long-form, as an overlay on the video for Shorts. Unrealistic, animated or slightly altered content is disclosed in the expanded description. Disclosure is required only for meaningful, photorealistic changes; the Help page excludes things such as beauty filters, colour adjustment, captions, upscaling, own-voice cloning, idea generation and production assistance such as outlines, scripts, thumbnails and titles. YouTube states the label alone does not change recommendation ...
• articleWhat are Gemini Omni Flash's documented capabilities, limits, prompting rules, watermarking and price in the Gemini API?
  Corrected 2026-10-24 - see C3-B978-01; the original statements below are kept for history and are superseded where the correction says so. Gemini API docs (captured 2026-10-23): gemini-omni-1.1-flash generates video with audio from text, image (reference/first frame), first+last frame interpolation and subject reference images, edits videos conversationally (multi-turn via the Interactions API) and extends videos at the tail. Output resolution 360p, 720p (default) or 1080p (upscaled); aspect 16:9 or 9:16. Limits: uploaded videos for editing/extension must be 10 s or less; extension appends to the end only; you cannot extend an uploaded talking clip to add dialogue (except via multi-turn on model-generated videos); voice editing and audio references unsupported; video references max 3 clips of up to 3 s each (audio ignored); no multi-video reasoning; no system instructions, temperature, top_p, stop sequences or negative-prompt field (write "Do not do X" in the prompt); YouTube URLs n...
• articleCan a filename_prefix with .., an absolute path or a symlink make ComfyUI write outside the output directory?
  Problem. Prefixes are user or agent text and could carry path-traversal sequences. Conditions. Core get_save_image_path at fb3befd. Method. After normalising the prefix, the code joins output_dir with the subfolder and calls is_within_directory, which resolves both paths with realpath and compares them with commonpath; if the folder is outside, it logs an error and raises, so the save fails and no file is written. Because realpath is used, a symlink inside the output directory that points elsewhere is also rejected at this check. Only the directory part is checked: the stem is the basename, so it cannot contain separators. An absolute prefix or '../' therefore should fail the check (inference from the code; not executed). Failure modes. The error aborts the node, so a pipeline that interpolates untrusted strings into the prefix gets a failed run rather than a rewritten path; sanitise first. The check happens at path computation, not at file open, so a symlink swapped in between chec...
• articleDoes YouTube say upload time affects recommendations or long-term performance, and for which formats does it say timing matters?
  Problem. Whether to optimise the clock time of a normal upload. Conditions. Stated for YouTube search and discovery generally. Method. Treat clock time as low-leverage for regular uploads: YouTube says publish time is not known to impact long-term performance and that recommendations aim to match the video to viewers regardless of upload date; treat timing as relevant for Live and Premiere, where viewers gather at one moment. Publishing when the audience is active may help early viewership, per YouTube, without long-term effect. Failure modes. Reading 'not known to impact' as 'proven no impact'; reading early-viewership benefit as a ranking boost. Limitations. The FAQ gives no numbers or experiment. Version/date. FAQ fetched 2026-10-24; undated.
• articleWhat does YouTube Help say the 'When your viewers are on YouTube' report covers (28-day window, purpose), and what does that page not state about time zone or thresholds?
  Problem. Choosing a Premiere or live-stream slot from audience activity. Conditions. YouTube Studio > Analytics > Audience. Method. Open the Audience tab; the report shows when viewers are online across YouTube (not only on your channel) over the last 28 days; YouTube names Premiere and live-stream planning and community building as uses. Failure modes. The fetched Audience page gives no time-zone basis, minimum-audience threshold or refresh cadence for this report, so do not assume the displayed hours are in your local zone or that small channels get a populated chart (inference, see E3). Limitations. Other Audience reports use 28-day or 7-day windows (monthly audience: previous 28 days from latest selected day). Version/date. Page fetched 2026-10-24; page carries no revision date.
• articleFor the open LivePortrait method, how should the driving video be prepared, and can I use it commercially?
  Problem: reuse a filmed face performance to animate a still portrait or video locally. Conditions: KwaiVGI/LivePortrait repo, inference.py with -s (source) and -d (driving) arguments. Method: crop the driving video to 1:1 (for example 512x512) or pass --flag_crop_driving_video; frame the head, minimise shoulder movement, and start on a frontal face with neutral expression. Transfer is facial and head motion only; it is a portrait animator, not a body or voice transfer, and it takes no audio input in the README's usage. Rights: the code is MIT licensed (Kuaishou), but the LICENSE file adds that InsightFace models are non-commercial research only and commercial users should replace InsightFace detection models. The README ethics note says synthesized results contain artifacts that may help detect deepfakes and disclaims legal responsibility. Failure modes: non-frontal starts and shoulder motion are warned against. Limitations: I did not run the code. Version/date: README fetched at st...
• articleDoes the cache stay portable across machines and repos: what do symlinks, shared Xet blobs and HF_HUB_DISABLE_SYMLINKS change, and is hf_transfer still used?
  Problem. A cache copied to a NAS, Windows host or second repo behaves oddly, or disk use is higher than expected. Conditions. Recent huggingface_hub with hf_xet. Method. The cache is symlink-based (snapshots/ points to blobs/); where symlinks are unsupported it stores files directly in snapshots/. HF_HUB_DISABLE_SYMLINKS=1 forces that mode, at the cost of duplicated big files, and the docs name a shared NAS used across operating systems as the use case. By default, Xet files are also stored once at <CACHE_DIR>/blobs/<prefix>/<xet_hash> and symlinked into repos so identical files across repos are not re-fetched; HF_HUB_DISABLE_SHARED_BLOBS=1 opts out, and it is off when Xet is disabled or symlinks are unsupported. Fast transfer is hf_xet (installed with huggingface_hub since 0.32.0); HF_HUB_ENABLE_HF_TRANSFER is deprecated, and HF_XET_HIGH_PERFORMANCE=1 uses more network and CPU. Failure modes. Older tools may delete a shared file used by other repos (it is re-downloaded); old cache ...
• articleA model repo is gated: what is the access-request flow, which token scope should a production worker use, and what does a failed fetch mean?
  Problem. Download of a gated model fails or a worker lacks access. Conditions. Repos with access requests enabled. Method. A human logs into their Hugging Face account, clicks Agree on the repo page (shares username and email, plus any form fields, with authors), and approval is automatic or manual per the authors. Scripts must then authenticate with a user token (hf auth login, login(), or HF_TOKEN). For production, a member requests access, then creates a fine-grained token with read access to that model, so the token cannot read other private repos; fine-grained tokens can also be shared in an organization. Failure modes. A rejected user cannot request again; authors can reset a decision, which revokes access and asks the user to agree again; huggingface_hub raises GatedRepoError (HTTP 403) when access is missing. Org policies may require approval or only fine-grained tokens (403 for read/write tokens). Limitations. Access is per user account and terms may restrict redistribution...
• articleWhich token does huggingface_hub use when several are available, and how do I stop a pipeline sending my token everywhere?
  Problem. A batch job authenticates with an unintended (e.g. write-scope) token or leaks it. Conditions. Scripts, containers, CI. Method. Order: a token passed as token= to a call; HF_TOKEN environment variable, which overrides the stored token; the token stored by hf auth login at $HF_HOME/token (or HF_TOKEN_PATH). Once logged in, all Hub requests, even ones needing no auth, send the token; set HF_HUB_DISABLE_IMPLICIT_TOKEN=1 to stop that. Use one token per app/machine so one can be revoked alone. Failure modes. Hardcoding the token in code or notebooks; reusing a write token for a read-only download. Limitations. Docs state the precedence of env over stored token; the exact rank of an explicit token= argument versus HF_TOKEN is not stated in these excerpts. Version/date: huggingface_hub docs main branch fetched 2026-10-24 (stamp per bibliography); behaviour may differ in older releases.
• articleHow do I pin a Hugging Face download to an exact commit so a generation pipeline reproduces, and why is a branch or tag name not enough?
  Problem. A workflow that downloads 'main' gets different weights after the repo owner pushes. Conditions. Any hf_hub_download, snapshot_download or hf download call. Method. Pass revision= (or hf://ID@REVISION) as a branch, tag, refs/pr/N or a FULL 40-character commit hash; record the hash that snapshot_download returns in the snapshot path. Short 7-character hashes are not accepted. A commit is immutable, so its file list is cached forever and a repeat download of the same commit needs no per-file network call. Failure modes. A branch or tag ref is re-resolved on each online run and refs/<branch> in the cache moves to the new commit; a mutable PR ref may change. Limitations. Pinning fixes which bytes you request, not whether they are safe to load (see C3-B710). Version/date: huggingface_hub docs main branch fetched 2026-10-24 (stamp per bibliography); behaviour may differ in older releases.
• articleIs a retention spike or dip good or bad, and can I tell skipping from quitting?
  Problem: editors over-read spikes as wins and dips as failures. Conditions: video meets eligibility (60 s, 100 views). Method: treat each marker as a prompt to inspect the footage at that timecode. YouTube lists two readings of a spike: viewers watched it more than earlier parts, or the content was unclear so viewers rewatched. A dip means skipped or stopped watching, and the page asks you to review why interest was lost. Also flat means watched start to finish, and gradual decline is normal tapering. Failure modes: assuming a spike means quality when it may mean confusion; assuming a dip means quitting when it may be skipping. Limitations: the report as described does not separate skip from exit or rewatch from share; segment views can exceed total views. Version/date: help page fetched 2026-10-24.
• articleWhat does the YouTube 'typical retention' comparison compare against: my 10 latest videos or all similar-length YouTube videos?
  Problem: benchmark claims about retention are ambiguous. Conditions: same Audience retention report. Method: the page names two baselines in different places: typical retention is described as comparing your 10 latest videos of similar length, while the SEE MORE view is described as comparing against all YouTube videos of similar length. Before quoting a benchmark, state which baseline the figure came from. Failure modes: calling a channel-relative figure a platform norm, or the reverse; comparing a long video to a short one (the baseline is length-matched). Limitations: the page does not give the length buckets, sample sizes or averaging method. Version/date: help page fetched 2026-10-24.
• articleWhy does my YouTube video show no key moments (intro, spikes, dips), and what minimum length, views and data delay apply?
  Problem: an agent or editor expects key-moment markers on every video and cannot tell missing data from a failed report. Conditions: Studio, video-level Analytics only. Method: check (1) video at least 60 seconds, (2) at least 100 views, (3) allow 1-2 days processing, (4) open the video's Analytics, Overview or Engagement tab, Audience retention report. Moments are highlighted only if detected, so any single type may be absent even on an eligible video. Failure modes: treating absence as proof of no spikes or dips; trying to read it at channel level; judging a Short under 60 s with this report. Limitations: the page states thresholds without saying whether views are all views or engaged views; the thresholds are the page's stated minimums, not guarantees of marker output. Version/date: help page fetched 2026-10-24, hl=en.
• articleHow does frontend pinning relate to workflow portability and to the frontend and untrusted-workflow security advisories?
  Problem. Choosing between the pip frontend, a pinned GitHub release and a local root when sharing workflows or hardening a machine. Conditions. Reproducing a workflow (C3-B746, C3-B590-05) or responding to frontend advisories (C3-B710-09/09). Method. The default path is a pip package with a version pin in requirements.txt, and ComfyUI logs a warning at startup if comfy* packages are below that requirement; updating means upgrading that pip package. --front-end-version overrides it per launch from GitHub releases and caches by folder; --front-end-root points to any local build and bypasses version management. The official docs list both flags with the same wording as the code. Failure modes. Because a failed pin falls back silently to the pip frontend (C3-B807-02), a portability note that says 'pinned to 1.x' should be checked against the running frontend, and a security fix delivered in a newer frontend is not applied by a pin to an older release. Limitations. I did not find a docs....
• articleWhere can hashtags be added on YouTube and YouTube Music, how are they suggested in Shorts, and where do hashtag links lead?
  Problem. Hashtag behaviour differs between videos, Shorts and playlists, and an agent needs the documented surfaces. Conditions. Uploading a video, recording a Short, or creating a YouTube Music playlist. Method. Add # in the title or description; the system suggests popular hashtags. For Shorts, suggestions come from signals such as past uploads and, in some countries, entered text; previously used suggestions carry a clock icon. Once saved, hashtags in titles and descriptions work as links to a results page of other videos and playlists sharing the hashtag. Of the hashtags in the description, up to three most engaging ones appear by the video title; the rest remain in the description. Failure modes. Which three show is chosen by YouTube, not the creator. Limitations. Page does not document landing-page ranking or ordering. Extends prior hashtag-count and pivot-page entries only with the Shorts and YouTube Music surface details. Version and date. Fetched 2026-10-24.
• articleBeyond the 60-hashtag and misleading-tag rules, which content-policy rules does YouTube apply to hashtags (harassment, hate, sexual, vulgar) and what are the stated consequences?
  Problem. Creators and agents drafting hashtags need to know that hashtags are policed like content, with different consequences by category. Conditions. Applies to hashtags in video or playlist titles and descriptions on YouTube and YouTube Music. Method. Check each hashtag against: harassment or hate (removal of video or playlist); sexual or explicit (may be removed); vulgar language (may lead to age restriction or removal); plus the non-hashtag rule (ordinary descriptive tags or repetitive sentences in the description remain prohibited). Policy-violating hashtags do not show below the content and may be removed. Failure modes. A single violating hashtag can affect the whole video or playlist, not just the tag. Limitations. The page gives no enforcement thresholds or appeal detail here; it points to separate policy pages. Version and date. YouTube Help page fetched 2026-10-24, no revision date shown.
• articleWhich security_level value ties Manager's high-risk features to the --listen address, and what exactly counts as safe? Docs say normal- blocks high-risk features only when --listen is set and the address does not start with 127.
  Problem. Choosing security_level for a ComfyUI-Manager install that may be reached over a network. Conditions. Manager config.ini security_level, ComfyUI started with or without --listen; docs page for Manager V3.38+ paths. Method. The page defines four levels: strong (no high or middle risk features), normal (no high; middle allowed), normal- (no high risk features only if --listen is specified with an address not starting 127.; middle allowed), weak (all). High risk = git-URL install, pip install, nodes outside the default channel, fix custom nodes. Middle = uninstall/update, default-channel installs, snapshot restore/remove, restart. Low = update ComfyUI. Reading the wording: normal- only differs from normal when --listen is given with a non-127. address, so a bare loopback run keeps high-risk features available while a LAN listen removes them. Failure modes. The test as worded is on the --listen argument text (inference: it names a prefix match on 127., not a check of who connec...
• articleWho can add a trailer to a YouTube Premiere, what must the trailer be, and which ads and fan-funding work during a Premiere?
  Problem. Adding a teaser and monetising the pre-start page. Conditions. Trailer: more than 1,000 subscribers and no Community Guideline strikes. Method. Upload the trailer as a normal video; in Studio schedule the main video as Premiere, Set up Premiere, then Add a trailer. Trailer must be 15 seconds to 3 minutes, any supported video type, ideally the same aspect ratio and resolution, non-infringing, guideline-compliant and not a previously premiered video. Ads: if eligible, pre-rolls only during the Premiere; mid-roll and post-roll are unavailable; Super Chat, Super Stickers and memberships (members-only chat) can be on before and during. Failure modes. Reusing a premiered video as the trailer is disallowed. Limitations. Ad eligibility defers to separate channel pages. Version/date. Fetched 2026-10-24.
• articleHow exactly do --listen, --port and the TLS flags behave, and is mutual TLS or any login offered?
  Problem. Exposing ComfyUI needs the right bind flags and the limits of the built-in TLS. Conditions. Direct exposure without a proxy. Method. --listen defaults to 127.0.0.1; bare --listen means 0.0.0.0,:: (all IPv4 and IPv6); a comma list binds several addresses on the one --port (default 8188), each started as its own TCP site. TLS turns on only when both --tls-keyfile and --tls-certfile are set (the help text says each requires the other; the code simply stays on http if only one is given, per the conditional, so check the startup log URL scheme). Failure modes. The SSL context uses verify_mode CERT_NONE, so no client-certificate authentication; there is no login, user or token check in these code paths (--multi-user only enables per-user storage). Port in use exits with an error. Limitations. Put authentication and rate limiting at a proxy or tunnel. Version. Commit fb3befd.
• articleWhat does --enable-cors-header actually do, and does enabling it weaken the built-in origin check?
  Problem. People add --enable-cors-header to let a web front end call ComfyUI. Conditions. Any start with the flag. Method. The flag takes one optional origin string (bare flag means '*'). Server.py chooses between two middlewares: if the flag is set it installs the CORS middleware, otherwise the origin-only middleware, so enabling CORS removes the Sec-Fetch-Site and Host/Origin 403 checks entirely. The CORS middleware answers every OPTIONS request with success and sets Allow-Origin to the configured string verbatim, Allow-Credentials true, and Allow-Headers including Authorization. Failure modes. No allowlist or per-request origin validation: a single static value; with '*' any page can drive the API (browsers refuse credentialed wildcard requests, but ComfyUI has no cookies to protect, so cross-origin simple requests can still queue work: inference). Limitations. Prefer a specific origin; put authentication in a proxy. Version. Commit fb3befd.

More