---
license: other
license_name: upstream-per-subset
license_link: https://huggingface.co/datasets/vanzl/DiffusionART-Prompts#sources-and-terms
task_categories:
- text-to-image
- text-to-video
- image-to-video
configs:
- config_name: pickscore
  data_files:
  - split: train
    path: pickscore/train.jsonl
  - split: validation
    path: pickscore/validation.jsonl
- config_name: t2v
  data_files:
  - split: train
    path: t2v/train.jsonl
  - split: validation
    path: t2v/validation.jsonl
- config_name: t2v1000
  data_files:
  - split: validation
    path: t2v1000/validation.jsonl
- config_name: i2v
  data_files:
  - split: train
    path: i2v/train.jsonl
- config_name: geneval
  data_files:
  - split: train
    path: geneval/train.jsonl
  - split: validation
    path: geneval/validation.jsonl
- config_name: ocr
  data_files:
  - split: train
    path: ocr/train.jsonl
  - split: validation
    path: ocr/validation.jsonl
---

# DiffusionART Prompts

Public prompt files used by the DiffusionART training infrastructure. Download
from this dataset; no upstream GitHub data download is needed at training time.

The converted JSONL files preserve source order and text. GenEval also preserves
the complete per-prompt scoring annotations. Existing PickScore/video manifests
match the previously pinned SHA256 values. GenEval/OCR are the named public
Flow-GRPO splits, not certified byte-identical copies of private historical runs.

| Configuration | Content |
| --- | --- |
| pickscore | Flow-GRPO train/test prompts |
| geneval | Flow-GRPO GenEval train/test prompts and annotations |
| ocr | Flow-GRPO OCR train/test prompts |
| t2v | DanceGRPO 50,000 training prompts and frozen VidProM validation100 |
| t2v1000 | Frozen VidProM validation1000, with validation100 as its prefix |
| i2v | DanceGRPO ConsisID 27,815 training prompts |

For I2V evaluation, use the t2v validation prompts. **No reference images are
included.** On-demand FLUX reference generation is a new evaluation protocol,
not a reconstruction of the original fixed JPEGs. All methods in a comparison
must share the reference images and evaluation settings.

## Usage

```python
from datasets import load_dataset
data = load_dataset("vanzl/DiffusionART-Prompts", "t2v", split="validation")
```

Training recipes pin an immutable dataset revision. `sources/` retains the
unchanged upstream inputs for deterministic conversion; `provenance.json`
records source URLs, hashes, row counts and converted-file hashes.

## Sources and Terms

- Flow-GRPO: https://github.com/yifan123/flow_grpo/tree/879042cf5707f8b90daa98d147d7deac2317c5da/dataset
- DanceGRPO: https://github.com/XueZeyue/DanceGRPO/tree/15cc71d53cc2e6e18a68ee607d5fb6ba9a99e344/assets
- VidProM: https://huggingface.co/datasets/WenhaoWang/VidProM (CC-BY-NC-4.0).
- ConsisID source: https://huggingface.co/datasets/BestWishYsh/ConsisID-preview-Data.

Upstream notices are retained in `attribution/`. This collection does not
relicense third-party data under the code repository's license. Respect each
source's terms, including VidProM's noncommercial restriction. FLUX-generated
images, when generated separately, are subject to the generator's output terms.

Paper: https://arxiv.org/abs/2610.12572
Code: https://github.com/vanzll/DiffusionART