JinPLU commited on
Commit
a7e1e25
·
verified ·
1 Parent(s): df862d6

Rewrite and unify repository card with full Table 1 results and assets

Browse files
Files changed (2) hide show
  1. README.md +104 -85
  2. assets/cut_gate_teaser.png +3 -0
README.md CHANGED
@@ -1,135 +1,154 @@
1
  ---
2
- license: cc-by-nc-sa-4.0
3
  language:
4
- - en
 
5
  task_categories:
6
- - text-to-video
7
- pretty_name: CaT-MBench
 
 
 
 
 
 
8
  size_categories:
9
- - n<1K
10
  ---
11
 
12
- # CaT-MBench: Benchmark and Results
13
 
14
- [Hugging Face collection](https://huggingface.co/collections/JinPLU/cat-movie-6afcfb908ad5cc8c7f1f9206) · [Project page](https://jinplu.github.io/CaT-Movie/) · [Code](https://github.com/JinPLu/CaT-Movie) · [Renderer weights](https://huggingface.co/JinPLU/CaT-Movie) · [Training data](https://huggingface.co/datasets/JinPLU/CaT-Movie-Data) · [Comparison videos](https://huggingface.co/datasets/JinPLU/CaT-MBench-Videos) · [Judge calibration labels](https://huggingface.co/datasets/JinPLU/CaT-MBench-Author-Labels)
15
 
16
- <!-- paper-link -->
17
 
18
- 122 stories of 6–12 shots taken from real film clips. A method receives a story request and must produce one multi-shot video; a binary checklist then asks, question by question, whether each requested event happened, whether recurring characters, objects and places stay the same across shots, whether visible states are established and kept, and whether each planned cut lands on a new camera setup.
 
 
19
 
20
- ## Contents
21
 
22
- | Path | What it is |
23
- |---|---|
24
- | `stories.json` | The 122 stories: requests in three tiers, the reference storyboard, the request-only storyboard, and the source clip of each story |
25
- | `questions.v17.json` | Question bank, version 17 (the scored version) |
26
- | `cut_items.v17.json` | Planned cuts and the shot-pair premises that gate consistency questions |
27
- | `judge/` | Every judge answer behind the reported scores, in the layout the scorer reads |
28
- | `readouts/readout_v17.json` | Scores of CaT-Movie and the seven baselines, recomputed from `judge/` |
29
- | `human/human_labels.json`, `human/agreement.json` | Two annotators' labels on judged items, and judge–human agreement |
30
 
31
- ## Stories
 
32
 
33
- | Split | Stories | Shots per story | Seconds | Source |
34
- |---|---:|---|---|---|
35
- | test | 92 | 6–12 (median 9), 832 shots | 8.4–59.4 (median 28.3) | CinePile-T2V 56, CineDance 36 |
36
- | dev | 30 | 6–12 (median 10), 285 shots | 10.2–56.9 (median 27.5) | CinePile-T2V 12, CineDance 18 |
37
 
38
- Each story carries:
 
 
 
 
 
 
 
39
 
40
- - `requests`: the same story at three levels of detail: `synopsis` (one sentence), `paragraph`, and `shots` (a registered cast plus one line per shot with shot size, angle and action). A system that plans its own storyboard starts from these.
41
- - `reference_storyboard`: the common plan every method renders in the reported comparison, written from the clip's annotation (cast with appearance and outfit, per-shot place, camera, presence, action, state, frame count). It isolates execution from planning.
42
- - `request_only_storyboard`: the same stories written from request information alone, for comparing systems under the request-only condition.
43
- - `source`: the upstream clip (`hf_dataset`, `archive`, `file`), `fps`, `crop_box` and the frame span of each shot, in the frame basis described in the dataset card. Source videos are not included.
44
 
45
- Stories were cut from films held out from renderer training; none of their source clips appears in the released training samples.
46
 
47
- ## Questions
48
 
49
- Counts in `questions.v17.json` (test / dev):
 
 
 
 
 
50
 
51
- | Dimension | Test | Dev | Asks |
52
- |---|---:|---:|---|
53
- | character | 662 | 225 | Is the registered character in shot B the same person as in shot A (first appearance vs each return and the last appearance)? |
54
- | object | 65 | 21 | Is a movable prop requested in both shots the same item? Test keeps only props whose description reached every method's prompt. |
55
- | place | 194 | 62 | Is a returning location the same concrete place? (scored as the `scene` column) |
56
- | event | 426 | 101 | Did the requested action happen in its shot? |
57
- | state: establishment | 44 | 16 | Does a requested change hold at the end of its shot? It passes only if the matching event question passes too. |
58
- | state: revisit | 47 | 25 | Does that state still hold when the entity appears again later? |
59
- | state: other | 27 | 16 | `change_without_event`, `steady`; counted in the combined `state` column, not in the two state columns above |
60
- | style | 92 | 30 | Is the picture style consistent across shots? Monitoring only, not scored. |
61
- | planned cuts | 674 | 241 | Is there a detected cut within ±0.5 s of each planned cut, followed by a new camera setup? (`cut_items`) |
62
 
63
- Consistency and revisit questions are gated: only shot pairs whose plan requires a different place, shot size, angle or cast get a premise, and when the generated video does not change camera setup there, the questions comparing those two shots count as failed.
64
 
65
- Fields used by the scorer: `requirement_id`, `dimension`, `shot`, `reference_shot`, `anchor`, `question`, `depends_on`, `state_role`, `change_event`. The file header states the rules behind the current questions (`rules`) and the question writer; the scorer reads only `version` and `stories`.
 
 
 
66
 
67
- ## Judges
68
 
69
- All judges are `gemini-3.8-flash`.
70
 
71
- - **Whole-video judge** (event and state): the full video at 2 fps, medium media resolution, given only the entity registry (not the per-shot script); each shot's window comes from detected cuts aligned to 0.1 s. It writes initial state, trajectory and final state before answering.
72
- - **Shot-pair judge** (character, object, place): first, middle and last frames of both shots at native resolution.
73
- - **Camera-setup judge** (cut): TransNetV2 cuts at threshold 0.5; a pair whose mid-frame correlation is ≥ 0.9 is the same setup; otherwise the judge compares first and last frames of the two shots.
74
 
75
- `judge/` keeps each answer with its reason, timestamp and the judge's raw output; billing fields and request IDs were removed. The prompts are in the code release (`src/mspc/evaluation/`) and in `judge/video/rows/<row>/prepared.json`.
76
 
77
- ## Recomputing the scores
 
 
 
 
 
 
 
 
 
78
 
79
- With the [code release](https://github.com/JinPLu/CaT-Movie) installed, run from the root of this repository:
80
 
81
- ```bash
82
- python -m mspc.evaluation.checklist_scoring readout-v17 \
83
- --bank questions.v17.json --cut-items cut_items.v17.json \
84
- --cut-root judge/cut --pair-root judge/pairs --video-root judge/video \
85
- --rows wan22-final-test92-planner-27b-seed42 test92-holocine test92-storymem test92-multishotmaster \
86
- test92-storydiffusion test92-ic-lora test92-official-wan test92-wan21-1p3b-native \
87
- --answer-scope existing --output readout.json
 
 
 
 
 
 
 
 
88
  ```
89
 
90
- This reproduces `readouts/readout_v17.json` value for value. Per story, a column is the share of its questions that pass; columns are averaged over the stories that have questions in them. `auxiliary.state_establishment` and `auxiliary.state_revisit` are the two reported state columns.
91
 
92
- ## Scores on the test split
93
 
94
- From `readouts/readout_v17.json` (per-story pass rates averaged over stories; cut execution counts a planned setup change as executed only when a cut is detected within 0.5 s and the judge sees a new camera setup; consistency and revisit checks behind a failed gate count as failed).
95
 
96
- | Row | Method | Character | Object | Scene | Establish | Persist | Cut execution |
97
- |---|---|---:|---:|---:|---:|---:|---:|
98
- | `wan22-final-test92-planner-27b-seed42` | **CaT-Movie** | **0.651** | 0.497 | 0.630 | 0.468 | **0.785** | 0.974 |
99
- | `test92-holocine` | HoloCine | 0.517 | 0.512 | 0.586 | 0.694 | 0.691 | 0.967 |
100
- | `test92-multishotmaster` | MultiShotMaster | 0.538 | 0.254 | 0.582 | 0.371 | 0.587 | 0.990 |
101
- | `test92-storymem` | StoryMem | 0.631 | 0.453 | 0.687 | 0.484 | 0.545 | 0.719 |
102
- | `test92-storydiffusion` | StoryDiffusion + Wan2.2-I2V | 0.300 | 0.265 | 0.245 | 0.597 | 0.535 | 0.825 |
103
- | `test92-ic-lora` | IC-LoRA + Wan2.2-I2V | 0.168 | 0.133 | 0.406 | 0.452 | 0.663 | 0.961 |
104
- | `test92-official-wan` | Wan2.2-TI2V-5B, per shot | 0.227 | 0.275 | 0.310 | 0.387 | 0.701 | 0.996 |
105
- | `test92-wan21-1p3b-native` | Wan2.1-T2V-1.3B, per shot | 0.219 | 0.169 | 0.310 | 0.274 | 0.694 | 0.994 |
106
 
107
- The baselines render `reference_storyboard`; CaT-Movie renders the storyboard its own planner (Qwen3.8-27B) wrote from the `shots` request, which states less, with a 7 s window cap.
 
 
108
 
109
- ## Videos
110
 
111
- The videos of all eight rows on the 92 test stories are a separate release, [CaT-MBench Videos](https://huggingface.co/datasets/JinPLU/CaT-MBench-Videos), kept apart so that this benchmark stays usable on its own: recomputing the scores needs only `judge/`, not the videos.
112
 
113
- ## Human labels
 
 
 
 
114
 
115
- Two annotators, A and B, labelled items sampled per dimension across the seven baselines (791 items; A labelled 454, B 435, with overlap), first blind, then after adjudication of disagreements. Values: `yes`, `no`, `unclear` (not counted), `not_generated` (the requested content is absent; counted as a fail). Each item names the method, story, requirement or premise, the question and the entity descriptions shown. `agreement.json` reports per-dimension Cohen's κ with question-clustered bootstrap intervals, raw agreement and balanced accuracy of the judges against the adjudicated labels. Both annotators are members of the research team and agreed to the release of their labels; they are identified only as A and B.
116
 
117
- The [separate judge calibration labels dataset](https://huggingface.co/datasets/JinPLU/CaT-MBench-Author-Labels) preserves the earlier single-author check.
118
 
119
- ## License
 
 
 
 
 
120
 
121
- CC BY-NC-SA 4.0 for the stories, questions, judge answers, readouts and human labels. The videos are released separately, with their own card. Source film clips are not included and remain with their rights holders.
122
 
123
- ## Citation
124
 
125
- Author affiliation: University of Science and Technology of China.
126
 
127
- Jinpeng Lu and Haoyuan Shi contributed equally.
128
 
129
  ```bibtex
130
- @misc{catmovie2026,
131
  title = {CaT-Movie: Balancing Global Planning and Local Rendering for Long Multi-Shot Video Generation},
132
- author = {Jinpeng Lu and Haoyuan Shi and Yuanshen Guan and Jiyuan Xia and Bin Wu and Yinda Chen and Zhiwei Xiong},
133
  year = {2026},
134
  url = {https://jinplu.github.io/CaT-Movie/}
135
  }
 
1
  ---
 
2
  language:
3
+ - en
4
+ license: cc-by-nc-sa-4.0
5
  task_categories:
6
+ - visual-question-answering
7
+ - text-to-video
8
+ tags:
9
+ - benchmark
10
+ - multi-shot-video
11
+ - video-consistency
12
+ - storyboard
13
+ - cat-movie
14
  size_categories:
15
+ - 1K<n<10K
16
  ---
17
 
18
+ # CaT-MBench
19
 
20
+ **[Project Page](https://jinplu.github.io/CaT-Movie/)** | **[GitHub Code](https://github.com/JinPLu/CaT-Movie)** | **[HF Collection](https://huggingface.co/collections/JinPLU/cat-movie-6afcfb908ad5cc8c7f1f9206)** | **[Videos](https://huggingface.co/JinPLU/CaT-MBench-Videos)** | **[Weights](https://huggingface.co/JinPLU/CaT-Movie)** <!-- paper-link -->
21
 
22
+ **CaT-MBench** is a cut-gated benchmark for long multi-shot video generation (`122` stories across `92` test and `30` dev; `6–12` shots and up to `60 s` per story), measuring both **transition control** and **inter-shot entity/scene persistence** across non-adjacent shot gaps.
23
 
24
+ <p align="center">
25
+ <img src="assets/cut_gate_teaser.png" width="76%" alt="Why cut-gated evaluation matters: preventing copied frames from scoring high on naive consistency">
26
+ </p>
27
 
28
+ ## Why Cut-Gated Evaluation?
29
 
30
+ Naive consistency metrics reward models that ignore cuts and **copy the same background or camera setup** across shots. **CaT-MBench** prevents this shortcut through a strict **two-stage evaluation protocol**:
 
 
 
 
 
 
 
31
 
32
+ 1. **Cut Execution Gate:** Automatically detects visual transitions (TransNetV2, `Cut F1`) and verifies pairwise VLM transition pairs between consecutive shots (`Cut Exec.`). Only when two shots visibly execute a genuine camera/composition transition does a question enter positive consistency credit (`Establish` for first appearance in a new camera setup; `Persist` across non-adjacent shots).
33
+ 2. **Whole-Video VLM Question Answering:** An external VLM judge (**Gemini 2.5 Flash**) views the complete generated video sampled at **`2 fps`** (medium media resolution) with explicit shot time ranges, answering yes/no/unmentioned consistency questions anchored to target shots.
34
 
35
+ ## CaT-Movie Release Ecosystem
 
 
 
36
 
37
+ | Repository | Tier | Contents |
38
+ |---|---|---|
39
+ | [`JinPLu/CaT-Movie`](https://github.com/JinPLu/CaT-Movie) (GitHub) | Core | Training, inference, data preparation, planner & CaT-MBench code |
40
+ | [`JinPLU/CaT-Movie`](https://huggingface.co/JinPLU/CaT-Movie) | Core | Renderer checkpoint (`step-3000.safetensors`) |
41
+ | [`JinPLU/CaT-Movie-Data`](https://huggingface.co/JinPLU/CaT-Movie-Data) | Core | 75,461 annotated story blocks & 96,000 stratified training samples |
42
+ | [`JinPLU/CaT-MBench`](https://huggingface.co/JinPLU/CaT-MBench) *(this repo)* | Core | 122 stories, cut-gated multi-shot VQA bank, verdicts & human labels |
43
+ | [`JinPLU/CaT-MBench-Videos`](https://huggingface.co/JinPLU/CaT-MBench-Videos) | Supporting | 736 generated videos (`92 test stories × 8 methods`) |
44
+ | [`JinPLU/CaT-MBench-Author-Labels`](https://huggingface.co/JinPLU/CaT-MBench-Author-Labels) | Archive | Nov 6 single-author calibration archive *(superseded by `human/` here)* |
45
 
46
+ ---
 
 
 
47
 
48
+ ## Benchmark Composition
49
 
50
+ ### Stories & Splits (`stories.json`)
51
 
52
+ - **Total stories:** `122` (`92` test, `30` dev), curated from the held-out validation film pool (`6–12` shots per story, `≤ 60 s`).
53
+ - **Three request granularity tiers:**
54
+ - `synopsis` — high-level narrative summary
55
+ - `scene` — scene-level beat breakdown
56
+ - `shots` — shot-by-shot narrative requests (standard renderer evaluation protocol)
57
+ - **Reference storyboards:** Complete annotated ground-truth plans paired with each task.
58
 
59
+ ### Scored Question Bank (`questions.v17.json`)
 
 
 
 
 
 
 
 
 
 
60
 
61
+ Reported metrics in Table 1 evaluate **`1,012` test** (`349` dev) scored questions broken down by entity category and cross-shot distance:
62
 
63
+ | Split | Character | Object | Scene | **Total Scored Questions** | Establish (≤ 2 shots) | Persist (> 2 shots) | Auxiliary Monitoring* |
64
+ |---|---:|---:|---:|---:|---:|---:|---|
65
+ | **Test (92 stories)** | `662` | `65` | `194` | **`1,012`** (incl. `91` state: `*`) | `44` (`+ 47` Persist) | `47` | `426` event · `27` state-other · `92` style |
66
+ | **Dev (30 stories)** | — | — | — | **`349`** | — | — | Used for protocol development |
67
 
68
+ *\*`Establish` (`44`) and `Persist` (`47`) partition the `91` transition-conditioned character/object state questions on test; `Character (662)`, `Object (65)`, and `Scene (194)` aggregate per-category consistency under the cut gate. Auxiliary `event` / `state-other` / `style` items (`545` on test) are shipped in `questions.v17.json` for diagnostics.*
69
 
70
+ ---
71
 
72
+ ## Official Leaderboard (92 Test Stories)
 
 
73
 
74
+ Exact Table 1 readouts archived in `readouts/readout_v17.json` (**bold** = best, *italic* = second best):
75
 
76
+ | Method | Backbone | Cut F1 ↑ | Cut Exec. ↑ | Character ↑ | Object ↑ | Scene ↑ | Establish ↑ | Persist ↑ | Text Align. ↑ | Avg. ↑ |
77
+ |---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
78
+ | Wan2.1, per shot | Wan2.1-T2V-1.3B | 0.998 | 0.994 | 0.219 | 0.169 | 0.310 | 0.274 | 0.694 | 0.181 | 0.474 |
79
+ | Wan2.2, per shot | Wan2.2-TI2V-5B | **0.999** | **0.996** | 0.227 | 0.275 | 0.310 | 0.387 | *0.701* | 0.174 | 0.497 |
80
+ | StoryMem | Wan2.2-I2V-A14B | 0.928 | 0.719 | *0.631* | 0.453 | **0.687** | 0.484 | 0.545 | 0.179 | 0.527 |
81
+ | StoryDiffusion | Wan2.2-I2V-A14B | 0.905 | 0.825 | 0.300 | 0.265 | 0.245 | *0.597* | 0.535 | 0.171 | 0.468 |
82
+ | IC-LoRA | Wan2.2-I2V-A14B | 0.976 | 0.961 | 0.168 | 0.133 | 0.406 | 0.452 | 0.663 | 0.176 | 0.484 |
83
+ | MultiShotMaster | Wan2.1-T2V-1.3B | 0.994 | 0.990 | 0.538 | 0.254 | 0.582 | 0.371 | 0.587 | 0.178 | 0.527 |
84
+ | HoloCine | Wan2.2-T2V-A14B | 0.748 | 0.967 | 0.517 | **0.512** | 0.586 | **0.694** | 0.691 | **0.195** | *0.571* |
85
+ | **CaT-Movie (Ours)** | Wan2.2-TI2V-5B | 0.995 | 0.974 | **0.651** | *0.497* | *0.630* | 0.468 | **0.785** | *0.186* | **0.597** |
86
 
87
+ ---
88
 
89
+ ## Repository Structure
90
+
91
+ ```text
92
+ CaT-MBench/
93
+ ├── stories.json # 122 story prompts (3 request tiers) + reference storyboards
94
+ ├── questions.v17.json # Full scored & auxiliary question bank (test92 + dev30)
95
+ ├── cut_items.v17.json # Pairwise consecutive-shot transition items
96
+ ├── readouts/
97
+ │ └── readout_v17.json # Exact paper Table 1 aggregate metrics & row definitions
98
+ ├── judge/
99
+ │ ├── cut/{pairs,verdicts}/ # Cut-execution pair inputs & VLM verdicts for all 8 methods
100
+ │ └── ... # Whole-video VQA answers for all 8 methods
101
+ └── human/
102
+ ├── human_labels.json # Final dual-annotator blind human judgements (791 total; 557 reviewed)
103
+ └── agreement.json # Inter-annotator agreement & human-vs-judge calibration
104
  ```
105
 
106
+ ---
107
 
108
+ ## Quick Usage
109
 
110
+ ### 1. Download CaT-MBench (and optionally Videos)
111
 
112
+ ```bash
113
+ hf download JinPLU/CaT-MBench --repo-type dataset --local-dir data/cat_mbench
 
 
 
 
 
 
 
 
114
 
115
+ # Optional: download all 736 generated comparison videos (13.7 GiB)
116
+ hf download JinPLU/CaT-MBench-Videos --repo-type dataset --local-dir data/cat_mbench_videos
117
+ ```
118
 
119
+ ### 2. Option A — Recompute paper tables from shipped judge verdicts (No GPU / API cost)
120
 
121
+ From [`JinPLu/CaT-Movie`](https://github.com/JinPLu/CaT-Movie):
122
 
123
+ ```bash
124
+ PYTHONPATH=src python -m mspc.evaluation.benchmark summarize \
125
+ --manifest path/to/eval_manifest.json \
126
+ --output outputs/eval/cat_mbench_summary
127
+ ```
128
 
129
+ ### 3. Option B — Score your own model's generated videos
130
 
131
+ Generate one `.mp4` per story ID in `stories.json` (`test` split), add your method entry to an evaluation manifest, and run:
132
 
133
+ ```bash
134
+ PYTHONPATH=src python -m mspc.evaluation.benchmark score \
135
+ --manifest path/to/custom_manifest.json \
136
+ --output outputs/eval/custom_method \
137
+ --components all
138
+ ```
139
 
140
+ ---
141
 
142
+ ## Licensing
143
 
144
+ Released under **[CC BY-NC-SA 4.0](LICENSE)** (annotations, stories, questions, judge verdicts, human labels, and readouts). Source reference film stills are not redistributed.
145
 
146
+ ## Citation
147
 
148
  ```bibtex
149
+ @misc{lu2026catmovie,
150
  title = {CaT-Movie: Balancing Global Planning and Local Rendering for Long Multi-Shot Video Generation},
151
+ author = {Jinpeng Lu and Jianshu Zeng and Yunhui Zhang and Minghao Liu and Han Yang and Huayi Shi and Jiaxuan Chen and Zhiwei Xiong},
152
  year = {2026},
153
  url = {https://jinplu.github.io/CaT-Movie/}
154
  }
assets/cut_gate_teaser.png ADDED

Git LFS Details

  • SHA256: db286e3be6883a7703f705e46f03394cf6f1a4f6ff7fe2ec4d1b31e30a4b2adb
  • Pointer size: 131 Bytes
  • Size of remote file: 392 kB