SEAM-Bench / README.md
Jackyqq's picture
Update README: HF download instructions and image-resize note
dac00f3 verified
|
Raw
History Blame Contribute Delete
8.06 kB
---
license: cc-by-nc-4.0
language:
- zh
- en
pretty_name: "SEAM-Bench: Continuity Storyboarding Benchmark"
tags:
- storyboarding
- short-drama
- visual-continuity
- multimodal
- film-generation
size_categories:
- n<1K
---
# SEAM-Bench
**A double-blind continuity storyboarding benchmark for industrial short-drama generation.**
SEAM-Bench standardizes the evaluation of *visual continuity* in AI short-drama
storyboarding and supports reproducible assessment at both the **prompt layer**
(storyboard text) and the **image layer** (rendered keyframes). It is the
benchmark released alongside **SEAM** (Shot Entity-Attribute Memory), a
training-free, model-agnostic memory graph that repairs cross-shot and
cross-episode continuity at the storyboard prompt-text layer, deployed as a stage
of the **SEAM-Agent** multi-agent storyboarding pipeline.
> SEAM-Bench 是面向工业级短剧生成的**双盲连续性分镜基准**,用于在提示词层
> (分镜文本)与图像层(渲染关键帧)上可复现地评测视觉连续性。
---
## Contents
The benchmark covers **three produced short dramas, 68 episodes in total**.
| Drama (dir) | Title | Episodes | Human storyboards | Reference images |
|---|---|---:|---:|---:|
| `beyond-the-wall` | Beyond the Wall | 20 | 659 | 31 (26 char + 5 scene) |
| `his-toyboy` | His Toyboy: The Billionaire's Trap | 23 | 754 | 15 (8 char + 7 scene) |
| `werewolf` | You Are My Cure, My Undoing | 25 | 826 | 12 (12 char) |
| **Total** | | **68** | **2,239** | **58** |
*Human-storyboard counts follow the aligned, annotated shot totals reported in
the paper; raw CSV row counts run higher because a single shot cell may span
multiple text lines.*
Two kinds of material are released:
1. **Original scripts** — per-episode screenplays, a tiered outline, and
character bios. These form the pipeline input.
2. **Reference images** — character and scene visual anchors, semantically
renamed and deduplicated into a shared pool. They serve as the fixed keyframe
input and as the image-layer consistency gold standard. For distribution the
images are resized to a 1536px long edge (PNG containers preserved); this is
ample for their use as reference anchors.
---
## Directory layout
```
dramas/
<drama>/
script/ # pipeline input (screenplays + metadata)
index.json # title, logline, per-episode scene/character index
tiered_outline.txt # multi-level story outline
character_bios.txt # character descriptions
camera_grammar.json # per-character shot-grammar rules (his-toyboy only)
episodes/
epNN_cn.txt # Chinese screenplay for episode NN
epNN_en.txt # English screenplay (his-toyboy only)
director/ # human-director storyboards, one CSV per episode
epNN.csv
character/ # character reference images
_pool/ # deduplicated unique images
reference_index.json # role/state -> image mapping, per episode
scene/ # scene reference images (absent for werewolf)
_pool/
scene_index.json
```
### `script/index.json`
```json
{
"title": "His toyboy: The Billionaire's Trap",
"logline": "...",
"num_episodes": 23,
"volumes": ["第一卷:...", "..."],
"episodes": {
"1": {
"title": "The Mark (咬痕)",
"summary": "...",
"scenes_en": ["EXT. MEDICAL UNIVERSITY COMMENCEMENT - STADIUM - DAY", "..."],
"characters_en": ["JULIAN", "VICTOR", "..."],
"scenes_cn": ["..."], "characters_cn": ["..."],
"num_lines_en": 72, "num_lines_cn": 69
}
}
}
```
### `director/epNN.csv`
Human-director storyboards. UTF-8 with BOM. Columns (Chinese headers):
| Column | 中文 | Meaning |
|---|---|---|
| 分镜号 / 集序号 | shot no. | Shot index within the episode |
| 场景 | scene | Scene / location label |
| 画面内容 | visual | Visual description of the shot |
| 景别 | shot size | e.g. 全景 / 中景 / 特写 |
| 拍摄角度 | angle | e.g. 平视 / 俯拍 / 仰拍 |
| 运镜 | movement | Camera movement |
| 角色 | characters | Characters active in the shot |
| 台词 | dialogue | On-screen dialogue |
| 中文台词 | dialogue (zh) | Chinese dialogue (where separated) |
> **Reading the CSVs:** cells may contain embedded newlines and commas, so parse
> with a proper CSV reader (e.g. Python `csv`, `pandas.read_csv`) rather than
> splitting on commas. Files are UTF-8 with a BOM — read with `encoding="utf-8-sig"`.
### `character/reference_index.json` and `scene/scene_index.json`
Map each per-episode reference to a deduplicated image in `_pool/`:
```json
{
"drama": "his-toyboy",
"pool_dir": "_pool",
"unique_images": 8,
"episodes": {
"ep01": [
{
"role": "朱利安",
"state": "julian_毕业典礼",
"image": "_pool/julian_毕业典礼.png",
"description": "朱利安,26岁,惊人漂亮但脸色苍白,毕业典礼上的毕业袍形象"
}
]
}
}
```
---
## Quick start
```python
import csv, json
from pathlib import Path
drama = Path("dramas/his-toyboy")
# 1. Story metadata
index = json.loads((drama / "script/index.json").read_text(encoding="utf-8"))
print(index["title"], index["num_episodes"])
# 2. A human-director storyboard (BOM-safe, multiline-cell-safe)
with open(drama / "director/ep01.csv", encoding="utf-8-sig", newline="") as f:
shots = list(csv.DictReader(f))
print(len(shots), "shots in episode 1")
# 3. Resolve a character reference image
refs = json.loads((drama / "character/reference_index.json").read_text(encoding="utf-8"))
for r in refs["episodes"]["ep01"]:
print(r["role"], "->", drama / r["image"])
```
---
## Notes and known gaps
Coverage is not uniform across the three dramas; the benchmark reflects the
material as produced.
- **English screenplays** (`episodes/*_en.txt`) exist only for `his-toyboy`;
`beyond-the-wall` and `werewolf` ship the Chinese screenplays only.
- **Scene reference pool** is absent for `werewolf` (no `scene/` directory).
- **`camera_grammar.json`** (per-character shot-grammar rules) is provided only
for `his-toyboy`.
- `werewolf/director/` additionally contains an aggregated
`... - 人类分镜总表.csv` (all-episode master table) alongside the per-episode
CSVs.
- Reference images (`character/_pool/`, `scene/_pool/`) are stored via Git LFS
on the Hugging Face Hub.
---
## Download
The benchmark is hosted as a Hugging Face dataset. Clone with `git` (LFS-backed)
or pull via the `huggingface_hub` client:
```bash
# Option A: git clone (install git-lfs first)
git lfs install
git clone https://huggingface.co/datasets/Jackyqq/SEAM-Bench
# Option B: Python client
pip install huggingface_hub
python -c "from huggingface_hub import snapshot_download; \
snapshot_download('Jackyqq/SEAM-Bench', repo_type='dataset', local_dir='SEAM-Bench')"
```
---
## Copyright and intended use
The screenplays, human-director storyboards, and reference images in this
benchmark originate from produced commercial short dramas and are released **for
non-commercial research on storyboard continuity evaluation only**. They remain
the property of their respective rights holders. Do not redistribute for
commercial purposes or use them to train generative models for commercial
release. If you are a rights holder and have concerns about any material here,
please open an issue.
> 本基准中的剧本、人类导演分镜表与参考图源自已制作的商业短剧,**仅供分镜连续性
> 评测的非商业研究使用**,版权归各自权利人所有。
---
## Citation
If you use SEAM-Bench, please cite the accompanying paper:
```bibtex
@inproceedings{seam2027,
title = {SEAM: Shot Entity-Attribute Memory for Consistent Short-Drama Storyboarding},
author = {},
booktitle = {Proceedings of the ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD)},
year = {2027}
}
```