CaT-MBench / README.md
JinPLU's picture
Eliminate Table 1 horizontal overflow, balance single-line badges, and brighten HF logo contrast
a425ad7 verified
|
Raw History Blame Contribute Delete
10.8 kB
metadata
language:
  - en
license: cc-by-nc-sa-4.0
task_categories:
  - visual-question-answering
  - text-to-video
tags:
  - benchmark
  - multi-shot-video
  - video-consistency
  - storyboard
  - cat-movie
size_categories:
  - 1K<n<10K

CaT-MBench

Project Page GitHub Code HF Collection Benchmark Videos Weights License

CaT-MBench is a cut-gated benchmark for long multi-shot video generation (122 disjoint held-out stories: 92 test + 30 dev; 6–12 shots, ≤ 60 s), evaluating transition control, inter-shot identity consistency, and state continuity (GPT-6.1-sol + compiled checks, Gemini 3.8 Flash verdicts & 557 reviewed blind human labels).

Overview of the Film Storyboard Dataset and CaT-MBench stories, GPT-6.1-sol question bank, and Gemini 3.8 Flash cut-gated judging

Why Cut-Gated Evaluation?

Naive cross-shot similarity (such as ViCLIP or DINOv2) rewards models that ignore planned cuts and copy the same background or camera setup. CaT-MBench removes this loophole with a two-stage protocol:

Cut-gated evaluation prevents frame copying from inflating consistency scores

  1. Transition Control (Cut F1 & Execution): Checks exact cut timing via TransNetV2 (Cut F1 @ 0.125 s across 740 planned test cuts) and setup changes across 674 cuts changing place, shot size, angle, or cast (Execution: TransNetV2 within ±0.5 s AND a new camera setup via mid-frame corr < 0.9 + 4-frame Gemini 3.8 Flash check). Any consistency or Persist check behind a failed cut gate receives 0 credit.
  2. Gated VLM Checks (Gemini 3.8 Flash) & Text Alignment: Character / Object / Scene compare 6 native-resolution frames per shot pair; Establish (change made in shot, paired with action) and Persist (state held on return) check the complete video at 2 fps; Text Alignment computes per-shot ViCLIP cosine similarity.
  3. Human Calibration (557 Reviewer-Adjudicated Blind Items): Across 7 baselines (94 shared pre-review overlap at κ = 0.64), Gemini 3.8 Flash attains κ = 0.704 ([0.64, 0.77], balanced accuracy 0.851 against pooled annotator blind κ = 0.853), vs. ViCLIP AUC 0.49–0.70.

Benchmark Composition

  • Stories (stories.json): 122 stories (92 test / 832 shots + 30 dev; 67 held-out films), each having three request tiers (synopsis, paragraph, shots) paired with annotated reference storyboards.
  • Question Bank (questions.v17.json): 1,012 test (349 dev) scored items (662 Character, 65 Object, 194 Scene, 44 Establish, 47 Persist) plus 545 auxiliary diagnostic items (426 event, 27 state-other, 92 style).

Official Leaderboard (92 Test Stories)

Exact Paper Table 1 readouts archived in readouts/readout_v17.json (seed 42; baselines render annotated reference storyboards in native prompt formats, whereas CaT-Movie renders its Qwen3.8-27B planner's storyboard from shots; bold = best, italic = second best):

Method (Video Base & Regime) Cut F1 ↑ Exec. ↑ Char. ↑ Obj. ↑ Scene ↑ Est. ↑ Persist ↑ Align. ↑ Avg. ↑
Wan2.1, per shot (1.3B, per-shot) 0.998 0.994 0.219 0.169 0.310 0.274 0.694 0.181 0.474
Wan2.2, per shot (5B, per-shot) 0.999 0.996 0.227 0.275 0.310 0.387 0.701 0.174 0.497
StoryMem (A14B-I2V, per-shot) 0.928 0.719 0.631 0.453 0.687 0.484 0.545 0.179 0.527
StoryDiffusion (A14B-I2V, per-shot) 0.905 0.825 0.300 0.265 0.245 0.597 0.535 0.171 0.468
IC-LoRA (A14B-I2V, per-shot) 0.976 0.961 0.168 0.133 0.406 0.452 0.663 0.176 0.484
MultiShotMaster (1.3B, joint) 0.994 0.990 0.538 0.254 0.582 0.371 0.587 0.178 0.527
HoloCine (A14B-T2V, joint) 0.748 0.967 0.517 0.512 0.586 0.694 0.691 0.195 0.571
CaT-Movie (Ours) (5B, joint) 0.995 0.974 0.651 0.497 0.630 0.468 0.785 0.186 0.597

(Bases: Wan2.1-T2V-1.3B, Wan2.2-TI2V-5B, Wan2.2-I2V-A14B, Wan2.2-T2V-A14B. Columns shorten Execution, Character, Object, Scene, Establish, Text Alignment.)

Quick Usage (Recompute Table 1 on CPU)

hf download JinPLU/CaT-MBench --repo-type dataset --local-dir data/cat_mbench

# From JinPLu/CaT-Movie: recompute exact Table 1 scores & paired bootstrap CIs
PYTHONPATH=src python -m mspc.evaluation.checklist_scoring readout-v17 \
  --bank data/cat_mbench/questions.v17.json --cut-items data/cat_mbench/cut_items.v17.json \
  --cut-root data/cat_mbench/judge/cut --pair-root data/cat_mbench/judge/pairs \
  --video-root data/cat_mbench/judge/video \
  --rows wan22-final-test92-planner-27b-seed42 test92-holocine test92-multishotmaster \
         test92-ic-lora test92-storydiffusion test92-storymem \
         test92-official-wan test92-wan21-1p3b-native \
  --answer-scope existing --output outputs/eval/readout_v17_recomputed.json

PYTHONPATH=src python -m mspc.evaluation.checklist_scoring paired \
  --readout outputs/eval/readout_v17_recomputed.json \
  --requests configs/benchmark/story_requests_v4.json \
  --pair test92-holocine wan22-final-test92-planner-27b-seed42 \
  --output outputs/eval/paired_vs_holocine.json
Score your own videos & Repository Structure

Download all 736 videos optionally via hf download JinPLU/CaT-MBench-Videos --repo-type dataset --local-dir data/cat_mbench_videos. To evaluate new .mp4 generations (gemini-3.8-flash):

# 1. Extract cuts: mspc.evaluation.checklist_scoring detect-cuts
# 2. Cut execution gate: prepare -> submit -> collect
PYTHONPATH=src python -m mspc.evaluation.cut_execution prepare|submit|collect ...
# 3. Shot-pair consistency (6 native-res frames) & whole-video states (2 fps)
PYTHONPATH=src python -m mspc.evaluation.consistency_pair_batch prepare|submit|collect ...
PYTHONPATH=src python -m mspc.evaluation.gemini_checklist_batch prepare|submit|collect ...
# 4. Aggregate under v17 bank
PYTHONPATH=src python -m mspc.evaluation.checklist_scoring readout-v17 --answer-scope complete ...
CaT-MBench/
├── stories.json                  # 122 story requests (synopsis / paragraph / shots) + reference storyboards
├── questions.v17.json            # Full scored & auxiliary question bank (test92 + dev30)
├── cut_items.v17.json            # Pairwise consecutive-shot & comparison transition items
├── readouts/readout_v17.json     # Exact Table 1 aggregate metrics & row definitions
├── judge/
│   ├── cut/{pairs,verdicts}/     # Cut-execution inputs & VLM verdicts for all 8 methods
│   ├── pairs/{pairs,receipts}/   # 6-frame shot-pair consistency inputs & judge receipts
│   └── video/rows/               # Whole-video 2 fps event/state answers & prepared metadata
└── human/
    ├── human_labels.json         # Blind dual-annotator judgements & reviewed test subset (557 calibrated items)
    └── agreement.json            # Pre-review overlap (κ=0.64), judge κ=0.704 & human calibration
CaT-Movie Release Ecosystem (GitHub & 8 HF Repositories)
Repository Tier Contents
JinPLu/CaT-Movie (GitHub) Core Planner, training, inference, curation & CaT-MBench code (Apache-2.0)
JinPLU/CaT-Movie Core Renderer weights (step-3000.safetensors)
JinPLU/CaT-Movie-Data Core 75,461 blocks, Qwen3.7-Flash schemas, grounding, 96,000 train windows & SAM 3.1 weights
JinPLU/CaT-Movie-T5-Cache Core Precomputed umT5-XXL BF16 caches (332,516 unique prompts)
JinPLU/CaT-Movie-VAE-Caches Core Precomputed VAE target latents (Wan2.2 704p + Wan2.1 480p)
JinPLU/CaT-Movie-Source-Videos Supporting Complete 18,173 block-yielding source MP4 clips
JinPLU/CaT-MBench (this repo) Core 122 stories, cut-gated GPT-6.1-sol VQA bank, Gemini 3.8 Flash verdicts & 557 human labels
JinPLU/CaT-MBench-Videos Supporting 736 generated test videos (92 test stories × 8 methods)
JinPLU/CaT-MBench-Author-Labels Archive Early Nov 6 single-author calibration archive (superseded by CaT-MBench/human/)

Licensing & Citation

Released under CC BY-NC-SA 4.0 strictly for non-commercial academic research. Referenced source video clips live in JinPLU/CaT-Movie-Source-Videos.

@misc{catmovie2026,
  title         = {CaT-Movie: Balancing Global Planning and Local Rendering for Long Multi-Shot Video Generation},
  author        = {Jinpeng Lu and Haoyuan Shi and Yuanshen Guan and Jiyuan Xia and Bin Wu and Yinda Chen and Zhiwei Xiong},
  year          = {2026},
  url           = {https://jinplu.github.io/CaT-Movie/}
}