Evaluation and media provenance

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

• articleIf I overwrite a checkpoint or LoRA file in place under the same filename, will ComfyUI's cache notice? Which core nodes hash file contents in IS_CHANGED?
  Problem. A long-lived headless server may keep serving cached results after a model file is replaced under the same name. Conditions. ComfyUI v0.42.5, default RAM-pressure or any non-null cache, same server process. Method. In nodes.py only three IS_CHANGED definitions exist (LoadLatent, LoadImage, and the channel-based LoadImageMask delegating to LoadImage); the first two hash the file with sha256. CheckpointLoaderSimple has no IS_CHANGED and its only input is the ckpt_name string chosen from folder_paths.get_filename_list. From the cache-key code (previous finding) the node signature therefore contains just the filename. Inference: replacing the file's bytes without renaming does not change the key, so a still-cached loader (and anything downstream) can be reused; use new filenames per version, or restart the server, or POST /free with free_memory true to reset the executor caches (main.py prompt_worker calls e.reset()). Failure modes. Provenance logs that record only the filename...
• articleHow is a CMOS (comparative MOS) test set up in a recent TTS paper, and what threshold is treated as no difference?
  Problem: decide whether two voice outputs differ audibly. Conditions: listeners compare a synthesized clip against a human or other-system clip for the same text. Method (as described by Seed-TTS): for each test sample evaluators first hear a reference clip of the target speaker; then hear the synthesized output and the ground-truth human recording in random order; they rate which has higher speaker similarity and expressiveness on a scale from -2 to +2 (the extremes mean least and strongest preference for the first sample); results are rearranged into one direction (system versus human) and averaged over evaluators and sentences. The authors state that an absolute CMOS below 0.1 is empirically considered insignificant between systems, and report CMOS of -0.07 (English) and -0.08 (Mandarin) versus human. Use (inference): randomise order and sample-to-listener assignment, hide system identity, include the human reference as a control, fix the reference clip, and report n listeners, n...
• articleWhat exactly counts as a YouTube thumbnail impression (the 1-second, 50%-visible rule), and which surfaces are excluded from impressions and CTR?
  Problem: CTR and impression numbers are only meaningful if you know what the denominator contains. Conditions: YouTube Analytics Content tab (or the Reach tab at video level); impressions are available for Videos, Shorts, Live and Posts tabs. Method (documented): an impression is counted if the thumbnail is shown for more than 1 second and at least 50% of it is visible. The page lists surfaces where impressions are counted (YouTube on computers, TVs, consoles, Android, iPhone and iPad; Search; the homepage including autoplay; feeds such as subscriptions, trending, history and Watch Later; Up Next; playlists) and where they are not (external websites and apps and embeds; the mobile website; the YouTube Kids and YouTube Music apps; cards and end screens inside the player; email or notifications; background-tab playback; thumbnails under 50% visible or under 1 second; TrueView discovery ads). Consequences: views from external embeds or notifications can exceed what impressions explain,...
• articleWhat guidance parameters does InstructPix2Pix expose for balancing the instruction against the original image, and what do FLUX.1 Kontext's paper and open-weight licence say for multi-turn edits?
  Problem. Instruction-based editing should change what you ask and keep the rest. Conditions. Local diffusers models. Method. InstructPix2Pix (Brooks, Holynski, Efros) edits an image from a written instruction in a forward pass with no per-example fine-tuning or inversion. The diffusers guide shows StableDiffusionInstructPix2PixPipeline with num_inference_steps=20, image_guidance_scale=1.5 and guidance_scale=10, and states that the two guidance scales control how much the original image and the edit instruction affect the result; training uses conditioning dropout so classifier-free guidance can be applied to the image and the prompt separately. Inference: raise image_guidance_scale to stay closer to the source frame, raise guidance_scale for stronger edits. FLUX.1 Kontext: the paper abstract states it handles local and global editing, character and style reference, and text editing in one model and reports improved preservation of objects and characters in iterative multi-turn workf...
• articleWhich lip-sync tools can be used commercially (LatentSync, MuseTalk, Wav2Lip, SyncNet) and where do dependencies break that?
  Problem. Code licence, checkpoint terms and dependency licences can differ. Conditions. Commercial or client delivery using open lip-sync tools. Findings from raw licence files and READMEs. LatentSync: LICENSE Apache 2.0; its data pipeline uses InsightFace landmarks (InsightFace pretrained models are non-commercial per its README, see C3-B571-06). MuseTalk: LICENSE MIT, README says no limit on academic or commercial use of the code, test data are non-commercial research only, and other open-source models used (whisper, ft-mse-vae, dwpose, S3FD) must comply with their licences. Wav2Lip: README prohibits commercial use (LRS2); no LICENSE file at the raw path. SyncNet python demo: MIT licence file read (Copyright 2016-present Joon Son Chung); the pretrained model's data terms were not checked. Method. Before use: (1) read each repo's LICENSE file, (2) list every downloaded checkpoint and its source page licence, (3) treat Wav2Lip as research only, (4) for MuseTalk and LatentSync confir...
• articleWhat does the Core Specification v1.0.1 normatively cover for USD, and what does it explicitly leave out?
  The AOUSD Core Specification 1.0.1 normatively covers USD's core: data types, document data model, paths, composition, stage population, value resolution, schema extension fields and file formats. It explicitly leaves domain uses (geometry, materials, physics) to other AOUSD specifications, keeps interpolation methods out of compliance validation and reserves fields such as expressionVariables and colorManagementSystem. The supplemental sample implementations are reference quality and not normative.
• articleCan an LLM be prompted to plan a suspenseful outline, and what did human evaluators say about the result?
  Partially verified: the curator re-read none of the cited sources; S9 rests on the producer's pinned reading. Xie and Riedl turn the 'diminishing escape routes' theory of suspense into a zero-shot prompting procedure: set up a protagonist, goal and dire stakes; then repeatedly ask for the protagonist's next action and a reason it fails, each less likely than the last; finally elaborate chapters. In pairwise human evaluations (90 raters per study), the method's stories were preferred over a plain 'write a suspenseful story' baseline for suspense (84.9% versus 11.8%), novelty and enjoyment, and the outline stage and detail stage were each needed; the same pattern held for Llama-2-13b-chat and versus Re3.
• articleDo open papers define standard metrics for branching story graphs (reconvergence, reuse, diversity), and what values do they report?
  Partially verified: the curator re-read none of the cited sources; S16, S17, S18 rest on the producer's pinned reading. In the three open arXiv papers read, no standard reconvergence or choice-density metric for authored branching stories is defined. MACHIAVELLI reports corpus scale only (134 games, 572,322 scenarios, 4,559 achievements). CHADPOD builds 1,462 decision-point classification tasks from CYOA graphs. Forking Garden (2026, LLM story-graph generator) defines an entity-overlap edge score and reports an entity-diversity ratio of 0.678 versus 0.261 for a baseline, a paper result on generated graphs, not on authored works.
• articleHow do I extract a graph from a finished branching work, what open graph data exists, and what choice-density measures have been used?
  Partially verified: the curator re-read none of the cited sources; S6, S7, S5, S8, S9 rest on the producer's pinned reading. Three first-person web sources: a 2017 post gives an edge-per-instruction extraction recipe (look at each page once, write an edge, label endings, render with Graphviz) and notes interest in reconvergence and cycles; the Transverse Reading Gallery states it holds 834 network graphs for 2,569 books with DOT and TGF files and edge lists; Swinehart's project classifies pages as branching, story or ending and plots page-type shares across 12 books, reporting a decline in endings and choices over time (qualitative reading of charts; no numeric table in the text read). Together they give an extraction method, open graph files and choice-density framing.
• articleIs there an open study of colour harmony judged on film and TV frames (multi-colour, real scenes), and what features did it find related to harmony?
  Partially verified: the curator re-read none of the cited sources; S7 rests on the producer's pinned reading. Wang et al. (Communication University of China, 2022; CC BY) had 84 non-expert college students aged 18-30 give 5-level harmony scores to 164 multi-colour materials taken from film and television screenshots after edge-smoothing with a 50-pixel circular mean filter, with a 16-dimension colour-feature set. Overall lightness, hue-tone differences, number of colours, lightness contrast, hue contrast and cool/warm contrast were correlated with harmony; lower contrast, brighter and cooler combinations tended to be rated harmonious. Internal consistency was reported (Cronbach's alpha 0.908), and a random forest classified harmony with 80.2% accuracy; the linear regression's best value was 63.9%. Coefficients and equations were not extractable from the HTML.
• articleWhat expert-designed tests can check creativity in a short story, how reliable are they, and how did LLM stories score?
  Partially verified: the curator re-read none of the cited sources; S6 rests on the producer's pinned reading. TTCW (Chakrabarty et al.) turns the Torrance dimensions (fluency, flexibility, originality, elaboration) into 14 binary yes/no tests proposed by 8 creative-writing experts, e.g. narrative pacing, scene vs summary, ending, originality in thought, character development. Ten new experts applied them to 48 stories (12 New Yorker stories plus 36 from GPT-3.5, GPT-4 and Claude v1.3 on the same plot, length-matched): New Yorker stories passed 84.7% of tests vs 8.7% (GPT-3.5), 27.9% (GPT-4), 30.0% (Claude v1.3). Per-test Fleiss kappa 0.27-0.66 (mean 0.41); aggregate pass counts correlate 0.69 between experts.
• articleDo automatic metrics (BLEU, ROUGE, BERTScore, BARTScore...) track human story judgments, and which should be used?
  Partially verified: the curator re-read none of the cited sources; S4 rests on the producer's pinned reading. On HANNA, 72 automatic metrics correlate mostly weakly (<30%) to moderately (30-50%) with human criteria at story level; system-level correlations are higher (40-70% for most) so metrics can rank systems but not individual stories. Best story-level metrics (|r| about 33-59%) include chrF, BERTScore, S3, BARTScore, SUPERT for relevance, and simple Repetition-3 and Novelty-1 statistics topped coherence and surprise; the authors advise relying on human annotation for ASG evaluation.
• articleHow does Re3 generate a 2,000-word story automatically, and what did its human evaluation and ablations show?
  Partially verified: the curator re-read none of the cited sources; S2 rests on the producer's pinned reading. Re3 (Yang et al., EMNLP 2022) has four modules: Plan (setting, characters, three-point outline via GPT3-Instruct-175B), Draft (recursively rebuilt prompt from relevant plan items, summaries and the last passage, GPT3-175B), Rewrite (rerank continuations with trained coherence and relevance classifiers plus heuristics) and Edit (character attribute contradiction detection and correction). Against GPT-3 rolling-window baselines, MTurk raters judged 14.3 and 20.0 points more Re3 stories coherent and premise-relevant; ablations show Plan and Rewrite matter and Edit does not; inter-annotator agreement was low.
• articleWhat does a first-person crew account of a splat-on-LED job look like, and does the best one found involve a world model?
  Partially verified: the curator re-read none of the cited sources; S9, S10 rest on the producer's pinned reading. The only first-person crew account found (DOSENMUTTER, Filmakademie, 2026-07-08) uses a captured photogrammetric splat, not a generated world; AI appears only as denoising. It is useful as a template of what is missing from the generative cases: stated alternatives tested (plates vs splat), a stated reason, capture counts and named roles. Berlin (Volinga case, reported by Radiance Fields) is likewise capture-based. Neither is evidence about world models.
• articleWhat checks or playtest measures do the open sources propose or use for player agency in interactive narrative?
  Partially verified: the curator re-read none of the cited sources; S7, S6, S3, S1, S4 rest on the producer's pinned reading. No inspected source supplies a validated agency instrument. What exists: Thue et al. asked 90 students to rate their experience after fixed or adaptive stories and tested whether adaptive stories felt more influential, with weak overall support and stronger results in filtered subgroups; Wardrip-Fruin et al. report Dow's study of three Facade versions using interviews, finding that greater presence and realism reduced agency by raising expectations; Facade's authors give self-estimated local-agency percentages; Mateas's structural model suggests an analytic check. Roberts and Isbell list measurability as a desideratum.
• articleWhich automatic metric does T2I-CompBench++ use for each compositional category (colour/shape/texture binding, 2D/3D spatial, numeracy, non-spatial, complex), and how well does each correlate with human ratings?
  Partially verified: the curator re-read none of the cited sources; S3, S4, S5 rest on the producer's pinned reading. T2I-CompBench++ (arXiv v3) has 8,000 prompts in four groups and eight sub-categories, and uses a different checker per sub-category: disentangled BLIP-VQA for colour/shape/texture binding, a UniDet detection (plus depth for 3D) rule for spatial relations and numeracy, CLIPScore or an MLLM for non-spatial relations, and a 3-in-1 average for complex prompts. Validated by Kendall tau / Spearman rho against 3-worker MTurk ratings, B-VQA is strongest for colour (tau 0.63) but weak for shape (0.27), UniDet is the best non-MLLM metric for spatial and numeracy, and plain CLIPScore is weak almost everywhere.
• articleHow well does GenEval agree with human raters, which settings were tuned to get that agreement, and where does it fail or stop discriminating?
  Partially verified: the curator re-read none of the cited sources; S1 rests on the producer's pinned reading. In the authors' MTurk study (6,000 annotations over 1,200 images) GenEval agreed with humans on 83% of images versus 88% pairwise annotator agreement and 80% for a per-task thresholded CLIPScore; on 860 unanimous images agreement was 91% vs 87%. The counting confidence threshold (0.9) and the colour crop+mask step were chosen against human labels, and the paper lists holes in objects, overlapping same-type objects and non-photographic art as failures. Single-object prompts are near ceiling for all tested Stable Diffusion/IF models, and the paper notes kappa 0 there because almost all images were correct.
• articleHow well have these benchmarks validated their automatic physics judges against humans, and can their validation numbers be compared?
  Partially verified: the curator re-read none of the cited sources; S1, S2, S4, S9, S12 rest on the producer's pinned reading. Five papers validate a judge or automatic scorer, each with a different statistic: VideoPhy-2-Autoeval Pearson 0.37 on physical commonsense (Gemini-2.0-Flash-Exp 0.11); PhyWorldBench's CAP ROC-AUC 75.1 for physical commonsense (plain GPT-o1 61.6); PhyGround's PhyJudge-9B aggregate relative bias 3.3 percent (Gemini-3.1-Pro 16.6) on held-out Veo-3.1 videos only; Last Exam's measurement evaluator 51.6 percent within-task pair agreement against 43.4 for a direct Qwen3.6-27B scorer and 56.6 between humans; Principia's VLMs 0.50-0.66 agreement with binary truth. The statistics are not comparable, none reports the same data for two judges, and all are the benchmark authors' own checks; the consistent message is that prompted general VLMs are weak physics judges and that fine-tuning or measurement helps modestly.
• articleHow does World Models' Last Exam in Physics test mechanics, optics, fluids and electromagnetism by measurement rather than by a VLM, and what do eight current video models score?
  Partially verified: the curator re-read none of the cited sources; S9, S11, S10 rest on the producer's pinned reading. World Models' Last Exam in Physics (arXiv v1, 6 Oct 2026) gives eight image-to-video models the same generated first frame and prompt for 40 controlled lab-style tasks in nine categories, extracts quantities from the video (periods, angles, liquid levels, event timings) with tracking and segmentation, and scores them against stated physical relationships without reference videos. A Qwen3.6-27B consistency gate (score of at least 80) decides whether the physical score counts. Seedance 2.5 averages 57.76/100, MiniMax H3 54.89 and CogVideoX1.5-5B 18.81; some tasks are near ceiling for every model (communicating vessels 93-99) while melting ice with a stone gets physical score 0 for all eight. Against 10 human annotators the evaluator beats a direct Qwen3.6-27B scorer (51.6 vs 43.4 percent pair agreement) but stays below human-human agreement (56.6).
• articleWhat do open-access studies report about readers reinterpreting earlier material after a twist (re-experiencing, hindsight), and which statements are measured versus proposed?
  Partially verified: the curator re-read none of the cited sources; S5, S6 rest on the producer's pinned reading. The one open-access measurement found is an fMRI study (PNAS 2026, 36 adults, one 18-minute audio story with a mid-story twist heard twice): participants rated a character differently on the second listen mostly in the pre-twist part, and passages they said they reevaluated showed larger neural pattern shifts. It is a mid-story twist, a single stimulus, and it measured reinterpretation, not liking of the ending. A 2026 theory paper proposes that a reversal pleases when earlier cues can be retrieved and reinterpreted; its support is three film close readings, and its three hypotheses are untested. Two further sources (Komeda et al., rereading twist stories; an open-access rereading article in Discourse Processes) could not be read.
• articleWhich quality claims in these 3D generation releases are measured results, which are vendor claims, and where do tables reuse or anonymise rows?
  Partially verified: the curator re-read none of the cited sources; S6, S29, S30, S34, S31, S28, S27 rest on the producer's pinned reading. All quantitative comparisons inspected were run and reported by the releasing teams; none is an independent evaluation. Three flags for curation: (1) the Hunyuan3D-2 README table lists "Top Open-source Model1" with numbers identical to the Trellis row in the Hunyuan 2.0 paper, and three unnamed closed-source rows; (2) Michelangelo, Craftsman 1.5 and Trellis rows in the Hunyuan 2.1 shape table are numerically identical to the same rows in the 2.0 paper, and the 2.1 texture table repeats the 2.0 paper Paint CLIP-FID, CMMD and CLIP values for "Hunyuan3D-2.0" but with a different LPIPS (0.1261 vs 0.0059); (3) the 2.1 title says "Production-Ready PBR Material" while its tables are image-level similarity metrics and contain no PBR-correctness or topology test. TripoSG ranks against methods with an LMM judge, and its VAE metrics are on its own data.
• articleWhen an agent passes HDR10+ or Dolby Vision metadata through hdr10plus_tool, dovi_tool and x265/FFmpeg, which steps are lossless, which alter or drop metadata by design, and which fail quietly?
  Partially verified: the curator re-read none of the cited sources; S12, S13, S10 rest on the producer's pinned reading. From the READMEs, docs and code at the pins (nothing executed): hdr10plus_tool inject needs a raw HEVC bitstream (extraction also accepts MKV), dovi_tool's HDR10+ SEI matching recognises only ST 2094-40 application version 1 HDR10+ SEI, and print but do not stop on a JSON/video frame-count mismatch (metadata is duplicated or skipped at the end). dovi_tool offers documented lossy operations (mode 2 removes profile-7 FEL mapping; editor remove_cmv4 deletes L3/L8/L9/L10/L11 and L254) and generating a Dolby RPU from HDR10+ yields only L1, so no L2/L8 colourist trims. FFmpeg's libx265 wrapper has no HDR10+ path while it enforces Dolby Vision metadata when configured; x265's own options carry both.
• articleWhat do libplacebo and FFmpeg's libplacebo filter actually consume from HDR10+ and Dolby Vision metadata when tone mapping, and what do they ignore or discard?
  Partially verified: the curator re-read none of the cited sources; S11, S10 rest on the producer's pinned reading. Per the libplacebo source at 13d0b798 (code reading only), HDR10+ scene statistics (MaxSCL, average maxRGB, the 99.99% distribution value) and, when present, the Bezier OOTF are mapped into its HDR metadata and feed peak/average-based tone mapping and the ST 2094-40 curve. For Dolby Vision it applies RPU reshaping and reads only Level 1 (max_pq, avg_pq) from the extension blocks; in the seven files scanned there is no use of L2/L8 trim values. Colourist trims are therefore not reproduced by this path, and the FFmpeg filter strips the COLOR_DEPENDENT side data when output transfer or primaries change.
• articleWhat documented failure modes affect Dolby Vision HDR-to-SDR metadata, and which can be checked before delivery?
  Partially verified: the curator re-read none of the cited sources; S2, S1, S3 rest on the producer's pinned reading. Netflix's QC section and Dolby's guide document a short list of Dolby Vision failure modes: L1 values of 0,0,0 on non-black shots (analysis done on the wrong active area or unusable), shot boundaries in the metadata that do not match cuts (mid-shot SDR shifts), wrong canvas/image aspect ratio (translucent bars, lifted matte blacks), positive lift above 0.025 on letterboxed content, cross-vendor tool incompatibility, and stale first-generation CMv2.9-compat blocks (CMVersion 4 0 instead of 4 1). Both sources name the Dolby Metafier validator, plus Dolby CM Offline for rendering the SDR, as the pre-delivery check.
• articleCan an automatic audio-visual sync detector or lip-sync metric (Synchformer offset classes, SyncNet-style LSE-D/LSE-C) certify that generated video and separately generated audio are in sync, and at what resolution?
  Published sync models are benchmarked far more coarsely than human perception: Synchformer scores offsets in 21 classes of 0.2 s and counts a prediction correct within +/-0.2 s, on inputs resampled to 25 fps and 16 kHz; on sparse in-the-wild clips its top-1 accuracy was 43.7% (60.2% within one class). Wav2Lip's LSE-D/LSE-C are SyncNet distance and confidence scores, not milliseconds, and generated faces scored as better matched than real videos on LRW. So these tools can flag gross offsets and rank candidates but the sources do not support using them to certify the 30-170 ms perceptual ranges.
• articleWhat do later open studies measure about LLM NPC dialogue in games (player ratings, grounding errors, prompt scaffolding), and how small or weak is the evidence for production writing?
  Two open studies were read in full-text excerpt form. Akoury et al. (Findings of EMNLP 2023) had 28 players compare GPT-4 infilled lines against Disco Elysium writers' lines and found designer prose preferred, with 32% of rated generations illogical in context. Figueiredo and Elumeze (arXiv 2510.25820v1) ran a 10-person usability study (no reliable difference after correction) plus a gpt-4o-judged synthetic probe of 60 scripted interactions in which scaffolding helped a quest-giver and hurt suspect NPCs. Both are small, single-game, short-horizon studies; neither measures lore consistency over long play.
• articleWhat exactly did the auditory Kuleshov experiment (Baranowski & Hecht) do and find about music and judged emotion of a face, what are its limits, and what do its stimuli rights look like?
  Read from the author's dissertation (Chapter 7, the 'in press' version; the journal body is paywalled and was not read, though the journal abstract matches). Thirty psychology students (21 female, 9 male, 20-31 y) watched 27 clips (3 per cell of a 3x3: music happy/sad/none x final-face expression neutral/happy/sad). Each 21 s clip: actor looking at a screen for 3 s, a 15 s muted neutral YouTube scene, 3 s final reaction. Music was unknown, instrumental, composer-categorised film music from online databases. Ratings: six basic emotions 0-5, with happiness and sadness analysed. Music affected happiness (F(2,58)=9.52, p<.001, partial eta squared .25) and sadness (F(2,58)=3.29, p=.044, .10) ratings; sad music lowered happiness ratings; happy music lowered sadness ratings versus sad music; sad music did not raise sadness significantly above no music; ratings had a floor effect. Stimuli rights (music databases, YouTube scenes) are not stated.
• articleHow does DPG-Bench score dense, long prompts, which judge does it use, and how was that judge validated against humans?
  DPG-Bench (introduced in the ELLA paper, arXiv v1) has 1,065 long prompts; GPT-4 wrote the dense prompts and, following the Davidsonian Scene Graph (DSG) recipe, per-prompt yes/no questions with dependencies. The released scorer asks an mPLUG VQA model each question on each of the four images in a 2x2 grid, counts an exact 'yes' as 1, zeroes questions whose parent question was answered 'no', and averages. Human validation described in the passages read is a ranking study of three models by 20 users per prompt, not a per-question agreement measurement of the VQA judge.
• articleWhat is VideoPhy-2, how are videos scored, how do systems score, and how far can its automatic evaluator be trusted?
  VideoPhy-2 (arXiv v1, Mar 2025) tests physical commonsense on 197 real-world actions with 3,940 LLM-written prompts. Three human annotators rate each video for semantic adherence and physical commonsense on 1-5, and the joint score counts videos with both at least 4. At v1 the best of seven models, Wan2.1-14B, reaches 32.6 percent overall and 21.9 on the hard split; momentum and mass conservation are the most violated laws (about 40 percent). Its 7B auto-rater improves on Gemini-2.0-Flash-Exp but physical-commonsense correlation is still only 0.37.
• articleHow does PhyGround score physics per law, how reliable is its human study, and how well does its open PhyJudge-9B judge track humans?
  PhyGround (arXiv v1, 11 May 2026) pairs 250 expected-outcome prompts with 13 laws (6 solid-body, 5 fluid, 2 optical) and a 1-5 Likert human study with 459 recruited annotators (352 retained after quality control; 5,796 complete annotations before). Eight models were scored: Wan2.2-27B-A14B and Veo-3.1 tie at 3.28/5 Overall and no model reaches 3.3; solid-body contact laws (momentum, collision) are hardest, and Veo-3.1 leads fluid and optical. The authors release PhyJudge-9B (a LoRA adapter on Qwen3.5-9B); its aggregate relative bias is 3.3 percent versus 16.6 for Gemini-3.1-Pro, measured only on held-out Veo-3.1 videos.
• articleWhat are the code, weights and dataset terms of the main 4D reconstruction and video-to-4D releases, and where do the texts conflict?
  Per-component record at pinned commits. Code: 4DGaussians (Apache text), Real-time 4DGS and DreamGaussian4D (MIT text), Consistent4D (Apache), Dynamic 3D Gaussians (MIT except Inria-licensed files) and SV4D inference code (MIT). Four repositories carry a conflict between the root licence and Inria non-commercial source headers or files (4DGaussians, Real-time 4DGS, DreamGaussian4D) or a stated split (Dynamic 3D Gaussians). Weights: SV4D 2.0 under the Stability AI Community License with a card that contradicts itself at exactly US $1,000,000 revenue; SV4D v1 card unreadable without login; Consistent4D and Dynamic 3D Gaussians advertise third-party or archive checkpoints without terms. Datasets: Neu3D is CC BY-NC 4.0; D-NeRF/HyperNeRF/Nerfies dataset terms are not stated in the READMEs read; DyCheck's iPhone dataset is Apache 2.0 per its README. No legal clearance is concluded.
• articleWhat did Park et al. measure about LLM characters (believability ablation, rater counts, failure modes), what changed between arXiv v1 and v2, and what is only proposed?
  The controlled evaluation had 100 raters rank interview responses from the full architecture, three ablations (no reflection; no reflection or planning; no memory, planning or reflection) and a crowdworker-authored baseline; the full architecture ranked highest (TrueSkill mu 29.89) and the memoryless ablation lowest. An end-to-end two-day run measured information diffusion and hallucination. Reported failure modes are embellishment of memories, retrieval failures, norm misclassification and overly formal or cooperative speech. A number differs between v1 (12 agents, 48%) and v2 (13 agents, 52%). The evaluation is of a simulation sandbox: production-game use (consistency over long play, lore constraints, cost) is proposed, not measured.
• articleFor each long-rollout video system, what code, weights and licence text is public at a pinned revision, and where do the code, card and file licences disagree?
  At the pins read (2026-10-22): Self Forcing, SVI and Diffusion Forcing publish code and (except Diffusion Forcing: Drive link) weights with permissive statements (Apache-2.0, MIT, MIT). CausVid (code/weights CC BY-NC-SA per files), LongLive (code Apache-2.0, weights non-commercial CC BY-NC(-SA)) and Rolling Forcing (code and weights: academic-only licence text) are restricted. Self-Forcing++ has no code, licence or weights at its pinned commit. Three conflicts exist between licence file, card prose and metadata.
• articleWhat do the HunyuanDiT, FlashVDM and bria-rmbg-1.4 agreements require of someone who reuses or redistributes TripoSG code or output: territory, user caps, output-use limits, attribution?
  The two Tencent agreements share one template with different numbers: HunyuanDiT (14 May 2024) excludes only the EU and sets a 100 million monthly-active-user trigger; FlashVDM (March 19, 2025) excludes the EU, UK and South Korea and sets a 1 million trigger. Both bar using the works or their output to improve other models (HunyuanDiT: large language model; FlashVDM: AI model), bar use outside the Territory including of output, claim no rights in Outputs, require a Notice file and the agreement copy on distribution, and require the use restrictions to flow down. The RMBG-1.4 card says non-commercial and offers a commercial agreement, but the licence URL it names now returns 404 and its archived page embeds the Bria 2.0 agreement, which does not name RMBG.
• articleWhat does MuSS report about keeping the same subject across separate shots when a reference image is given, and why do reference-conditioned models look good on reference fidelity but fail between shots?
  MuSS (arXiv 2604.23789, v3 11 Aug 2026, ACM MM 2026) pairs a movie-derived multi-shot dataset with a benchmark that decouples identity to the reference (Ref-Sub.Con) from identity between generated shots (Inter-Sub.Con, DINOv2 and InsightFace) and adds ACP-Var and CP-Rate for copy-paste. On 50 Track-2 test items the authors report Phantom 75.16 / 55.20 and VACE-14B 72.10 / 52.40 (Ref / Inter) versus their own EchoShot-based fine-tune 78.50 / 62.27, with CP-Rate 18.50% (Phantom), 21.68% (VACE-14B) and 7.35% (theirs). Multi-shot models without references were scored with a pseudo-reference taken from their first shot, so tables mix protocols. Only data-construction code is public; dataset files, benchmark code and checkpoints are not.
• articleFor each Kuleshov experiment (Mobbs 2006, Barratt 2016, Calbi 2017, Mullennix 2019, Cao 2024, Szaszko & Loebus 2026), what were the subjects, stimuli, timings and rating labels, so a protocol can be compared or rebuilt?
  A side-by-side methods map from the papers' own text. Samples are small and mostly student or volunteer (14, 28, 32, 36, 44, 59 plus 31 fMRI). Sequence order differs (Mobbs, Mullennix and Szaszko & Loebus put context before the face; Calbi, Cao and Barratt use face-scene-face). Timing ranges from 0.75 s faces with 4-8 s gaps (Mobbs) to self-paced viewing (Mullennix, Szaszko & Loebus). Rating labels differ by study (Russell-circumplex 2-D scale, SAM-derived visual analogue scales, 9-point valence/arousal, 5-, 7-label category lists), which is the main reason results cannot be tallied across studies. Barratt 2016 could only be described from abstract plus three later papers' descriptions, which disagree on the label count.
• articleWhy do latent action models learn noise, camera motion or scene cues instead of actions, and which remedies do the papers report?
  Four papers analyse why latents go wrong. Zhang et al. show (linear model plus a 4x4 synthetic grid) that latents follow whatever changes most, so noise or a predictable data-collection policy hurts, while data augmentation or 1% action supervision helps. Lee et al. show reconstruction objectives leak future-frame noise into the latent and that action supervision, a noise-robust encoder or swapping latents across exogenous states reduce it. Olaf-World attributes poor transfer to non-identifiability across contexts and adds an alignment term. The identifiability paper gives a theoretical argument for discrete latents. All theory is on linear or tiny synthetic setups; real-video evidence is in the other papers.
• articleHow do the primary papers evaluate latent actions, and which of those checks could an agent re-use on its own model?
  Every paper uses a different test, so numbers cannot be compared across papers. Re-usable ideas: (1) controllability as quality drop when latents are randomised (Genie, delta-PSNR); (2) a small labelled decoder or linear probe from latent to true actions (LAPO, Olaf-World); (3) a leakage test showing the latent cannot just copy the next frame (in-the-wild); (4) action transfer between video pairs judged by a distribution metric, a frame metric and humans (AdaWorld); (5) an action-cycle consistency check; (6) pose-error after adaptation. Most real-video transfer judging is qualitative or small human panels.
• articleWhen an AI tool "rotos" a subject, is the result a segmentation mask or an alpha matte, and which metrics (J&F versus MAD/MSE/Grad/Conn/dtSSD) apply to each?
  Video-object-segmentation models such as SAM 2 predict binary object masks and are scored with J&F, which in the DAVIS protocol is a region IoU plus a boundary F-measure on binary masks. Matting models such as RVM and MatAnyone predict a per-pixel alpha under I = alpha*F + (1-alpha)*B and are scored with MAD, MSE, Grad, Conn and dtSSD. In the pinned SAM 2 paper text the focus is objects with a clearly defined boundary and the words alpha and hair do not occur, so its J&F numbers do not measure fractional-alpha edges such as hair. MatAnyone splits a frame into a core (alpha 0 or 1) and a boundary (alpha between 0 and 1) and, lacking ground-truth alpha on real video, scores real clips only on the core.
• articleHow can I produce one comparable distinctiveness score per character, and what does the score actually measure?
  Seła et al. (arXiv v1, 2023) define a character's distinctiveness as the degree to which their speech differs from the other characters in the same play and measure it two ways: median bootstrap energy distance between the character's character-3-gram distribution and everyone else's (statistically stronger), and the area under the curve of a character's 100 top weighted-log-odds keywords (more interpretable). Applied to 3301 characters in 2324 works (DraCor: French, German, Russian, Shakespeare), the score is largely driven by dialect, accent, class markers and topic, and falls as a character's share of dialogue rises. It is a within-play statistic, not a quality score.
• articleMy draft's characters all sound the same. What simple check can I run to confirm it and find which lines to fix?
  Baker (1919) gives two manual checks that need no software: swap each speaker label for a bare type name ('A man', 'A young woman') and see whether the dialogue still identifies who is speaking, and ask of each speech whether the imagined speaker, not the author, would say it. The computational papers add automatic versions of the first idea: if a classifier cannot tell speakers apart from their lines, or character-vs-rest distance is near the self-baseline, the voices are not distinct. No source gives a pass mark for a draft.
• articleDoes any open screenplay-corpus paper report scene lengths or act proportions in minutes?
  STAGE (arXiv 2601.08510v2) reports screenplay length in words and structure in scene counts per language (English 108 screenplays: mean 12,453 words, 137 scenes) and does not convert to minutes; its released artefacts are derived-only and non-commercial.
• articleHow fast is typical cutting in Hollywood features and how has average shot length changed, as a benchmark for pacing a shot list?
  Cutting, Brunick, DeLong, Iricinschi and Candan (2012) report for 160 English-language films 1935-2010 (high box-office or, before 1977, most-rated on IMDb) that average shot length fell from about 10 s in the 1930s-40s to below 4 s after 2000 (r = -.75, all five genres); the attention-control explanation is their hypothesis.
• articleWhat do open-access papers that map face-tracker blendshapes to FACS action units report about methods and limits?
  A CC BY Sensors 2026 study (PMC12899966) maps ARKit blend shapes to action units using a published guide (its ref. 11) and reports 68.3% seven-class posed-emotion accuracy against a 58.9% mean of three human raters, while stating that ARKit coefficients do not always map cleanly to FACS units.
• articleHow does Smithsonian Open Access define which records are CC0, what does its API require, and what could not be read about its licence fields?
  The Smithsonian Open Access FAQ, as archived on 12 Sep 2025, treats CC0 as a per-item designation and says items not designated CC0 need prior written permission for commercial use; metadata access is via a keyed api.data.gov API.
• articleWhat do film emotional-arc studies (Del Vecchio et al.; Chu and Roy) measure and report, with what data, methods and stated limits, so an agent can say what a computational "story arc" does and does not show?
  Del Vecchio et al.'s preprint (arXiv 1807.02221v2, July 2018) clusters lexicon sentiment arcs of 6,174 English subtitle files from opensubtitles.org (its abstract calls them 'movie scripts') into six shapes and reports Man in a Hole as the highest mean domestic gross ($37.48 million), which the authors attribute to being 'talked about' rather than 'liked'; cluster-count robustness is not in the paper. Chu and Roy (arXiv 1712.02896v1) found combined audio-visual arcs on 509 films formed largely indistinct clusters.
• articleDo measured shot-duration patterns across 150 popular films support a 30/60/30 three-act split, and what act timings does the empirical literature actually report?
  Cutting (2016, CC BY 4.0) reports, for a 150-film set of English-language popular films (a 160-film corpus 1935-2010 culled of films over 2.5 h and with intermissions), shot-transition and luminance patterns he reads as better fitting Thompson's four-act scheme than Field's three-act scheme: about 98% of protagonists appear in the setup window and mean luminance is lowest near the three-quarter point. Act proportions (about a quarter each) are Thompson's parsing of 73 films, quoted, not re-derived.
• articleWhat do Tencent's Hunyuan3D-2 licence and README say about territory, user limits, output rights and PBR, and does this block studio use?
  Problem. A model labelled open-source may carry territorial and user-count conditions. Conditions. Using Hunyuan3D-2 family weights or code. Method. (1) Read the Community License: it is limited to a Territory that excludes the EU, UK and South Korea; use of works or Output outside it is unlicensed. (2) If the licensee's products exceed 1 million monthly active users on the release date, a licence must be requested from Tencent. (3) You must not use the works or Output to improve another AI model; Outputs themselves are not Model Derivatives and Tencent claims no rights in Outputs, but responsibility for Outputs sits with the user. (4) Pipeline: README describes a two-stage design (bare mesh, then texture synthesis) and VRAM of about 6 GB for shape and 16 GB for shape plus texture; the 2.1 release is announced as adding a PBR model, under its own repo and licence (not read here). Failure modes. Using outputs for a European release; training a downstream model on outputs. Limitations...
• articleWhat does TRELLIS output for an image-to-3D asset, and what do its README and licence imply for game or film use?
  Problem. Image-to-3D tools produce meshes of unclear production readiness. Conditions. Using Microsoft TRELLIS to make assets from an image. Method. (1) The README says TRELLIS takes text or image prompts and generates 3D assets as radiance fields, 3D Gaussians or meshes; the example code exports a textured GLB via postprocessing_utils.to_glb with a simplify ratio (0.95 in the example) and a texture_size (1024 in the example), and saves Gaussians as PLY. (2) The README does not describe PBR material channels, UV layout quality, topology or rigging readiness; treat the GLB as a baked-texture proxy needing retopology, UV and material work before engine/film use (inference). (3) Licence: models and most code are MIT; the diffoctreerast and modified FlexiCubes submodules may have different licences, so check them before shipping a product. Failure modes. Assuming MIT covers the whole tree. Limitations. No visual inspection of outputs was performed. Training-data licence questions are co...
• articleCan I use the Inria reference 3D Gaussian Splatting code in a commercial or client project, and which popular alternatives carry different licences?
  Problem. Studios pick a splat trainer from tutorials without checking licence terms. Conditions. Any paid, client or product use of splat training or rendering code. Method. (1) Read LICENSE.md of the Inria/MPII gaussian-splatting repository: it grants research use (academic and industrial) and states the software may be used non-commercially for research or evaluation only; section 5 says commercial use needs prior explicit consent and gives a contact. (2) Treat derivative works as bound by the same use limitation (section 4.2). (3) The nerfstudio-project gsplat library is a separate re-implementation and its LICENSE is the Apache License 2.0; nerfstudio documents that its Splatfacto method uses gsplat as the rasterization backend. (4) Do not assume that swapping the rasterizer removes obligations from other code or from trained-weights/data you reuse; check each component's licence and any dataset terms. Example. A pipeline that shells out to the Inria train.py inherits the non-co...

More