ShotBench: expert-level cinematic understanding benchmark for VLMs
Source: https://arxiv.org/abs/2506.21356
L07 2026-10-17: merged duplicate records from x03:arxiv-2506-21356-shotbench (same canonical URL).
Structured fields
{
"checked_at": "2026-10-17T00:00:00.000Z",
"usefulness": "3.5k+ expert QA pairs from 200+ acclaimed films across eight cinematography dimensions; measures whether VLMs can name shots, lighting and camera work correctly.",
"source_type": "paper",
"availability": "available",
"source_owner": "Liu et al. (Vchitect)",
"next_check_at": "2026-10-31T00:00:00.000Z",
"known_limitations": [
"The eight dimensions are not named on the abstract page; read the paper before citing them."
],
"upstream_updated_at": "2025-06-27T00:00:00.000Z"
}Evidence claims
- [sourced_fact] Abstract: over 3.5k expert-annotated QA pairs from images and clips, curated from 200+ films (predominantly Oscar-nominated), spanning eight dimensions. · locator: abstract source
- [upstream_claim] Abstract: evaluation of 24 leading VLMs reveals substantial limitations. · locator: abstract source
- [sourced_fact] Repository Vchitect/ShotBench has no license file per GitHub; last push 2025-09-12. · locator: repository metadata source
- [upstream_claim] Over 3,500 expert-annotated QA pairs from 200+ films spanning eight cinematography dimensions; 24 models tested; best under 60% average accuracy; ShotQA (~70K pairs) and ShotVL released. · locator: abstract source
- [sourced_fact] Accepted at NeurIPS 2025; CC BY 4.0 on arXiv. · locator: comments and license source
- [upstream_claim] Research review verdict at export: reviewed (as of 2026-10-17). Recorded by the research pipeline, not a hub review outcome; evidence.review stays unreviewed until an independent reviewer records one.
Rights
| component | license | hosting | redistribution | modification | training | Evidence and notes |
|---|---|---|---|---|---|---|
| documentation | CC-BY-4.0 | permitted | permitted | permitted | unknown | arXiv page lists CC BY 4.0 for the paper. |
| code | unknown | unknown | unknown | unknown | unknown | No license file detected. |
| dataset | unknown | unknown | unknown | unknown | unknown | Benchmark content is drawn from films; dataset license not checked. |
Linked records
- How reliably do current video generators follow named camera moves and lens language in text prompts? discusses → this
- Do expert-annotated cinematography benchmarks predict usefulness for real shot planning? discusses → this
- Does giving a planning agent a cited film-style study measurably change its shot list versus a control? discusses → this
- Can evaluation keep continuity, composition, lighting intent, camera feasibility, aesthetic preference and narrative purpose as separate dimensions without collapsing into a leaderboard? discusses → this