---
license: apache-2.0
configs:
- config_name: default
  data_files:
  - split: train
    path: metadata.csv
---

# CutCraft

CutCraft is a benchmark dataset for **professional editing-technique execution in multi-shot audio-video generation** (*Beyond Coherence: Benchmarking Professional Editing-Technique Execution in Multi-Shot Audio-Video Generation*).

It contains **4,720 generated videos** produced by **16 video generation/editing backends** (295 benchmark prompts each). The videos of each backend are packed in a single zip archive.

## Data structure

Each zip archive `{model}.zip` contains 295 videos named `{video_id}_{title}.mp4`.

The `metadata.csv` file (shown in the Dataset Viewer) lists all 4,720 videos with the following columns:

| column | description |
|---|---|
| `model` | name of the generation/editing backend (same as its zip archive name) |
| `video_id` | benchmark prompt ID |
| `title` | short title of the prompt |
| `filename` | video file name inside the zip archive |
| `zip_file` | zip archive containing the video |
| `zip_url` | direct download URL of the zip archive |

## Backends

Agent_happyhorse, Agent_ltx2, Agent_wan2.7, Davinci, JavisDiT++, MiniMax-H3, happyhorse, kling, ltx2, mova, ovi, seedance2.0, seedance2.5, veo3.1, vidu, wan2.7

## Paper

For more details, see the paper on [arXiv:2609.08275](https://arxiv.org/abs/2609.08275).