MGFlow-T2I

Synthetic image-prompt pairs used as the reference distribution for MGFlow text-to-image post-training in Unifying Distributional Training for One-Step Visual Generation. The same reference images support both image-only and joint image-text matching.

Project page: https://shihaoyang0423.github.io/MGFlow-website

Model checkpoints and fitted reference distributions: shy0423/MGFlow.

Dataset contents

The dataset contains 301,706 synthetic images at 512×512 with English prompts. All images were generated with the official distilled FLUX.2 [klein] 4B model using four sampling steps and guidance scale 1.

Subset	Prompt source	Images	TAR shards
COCO	82,783 COCO train2014 captions	248,349	105
GenEval	553 GenEval prompts	53,357	19
Total		301,706	124

Reference construction follows the procedure used by iRDM and AMFD:

• COCO: use the first caption in annotation order for each train2014 image, generate 24 candidates per caption, and retain the three with the highest PickScore.
• GenEval: retain up to 100 candidates per prompt that pass the official GenEval correctness test, considering up to 1,000 candidates when needed. Some prompts yield fewer than 100 passing images.

The release contains generated images, not original COCO photographs. See the paper's text-to-image training details for the full construction protocol, and ATTRIBUTION.md for image and prompt sources.

File structure

data/coco/train-*.tar       COCO-derived image-prompt pairs
data/geneval/train-*.tar    GenEval image-prompt pairs
metadata.jsonl             One record per image: id, prompt, subset, shard, image
manifest.json              Image counts, shard sizes, and checksums
ATTRIBUTION.md             Source attribution and license notices

Each TAR shard stores pairs named <id>.png and <id>.json in WebDataset format. metadata.jsonl maps every image ID to its prompt and location in a shard; manifest.json lists the per-shard counts and SHA-256 checksums.

Read a sample

Install huggingface_hub and pillow. This example streams the first metadata record and its image without downloading the full dataset or a complete shard:

import json
import tarfile
from urllib.request import urlopen

from huggingface_hub import hf_hub_url
from PIL import Image

repo_id = "shy0423/MGFlow-T2I"
metadata_url = hf_hub_url(repo_id, "metadata.jsonl", repo_type="dataset")
with urlopen(metadata_url) as response:
    sample = json.loads(response.readline())

shard_url = hf_hub_url(repo_id, sample["shard"], repo_type="dataset")
with urlopen(shard_url) as response:
    with tarfile.open(fileobj=response, mode="r|") as archive:
        for member in archive:
            if member.name == sample["image"]:
                with archive.extractfile(member) as image_file:
                    image = Image.open(image_file).convert("RGB")
                break

print(sample["prompt"])
print(image.size)

License

Generated images and MGFlow-authored metadata are released under CC BY 4.0. Third-party prompts retain their original licenses; see LICENSE.md and ATTRIBUTION.md.

Citation

If you use this dataset, please cite:

@misc{zhang2026unifyingdistributionaltrainingonestep,
  title={Unifying Distributional Training for One-Step Visual Generation},
  author={Chi Zhang and Shi Haoyang and Yueyi Liu and Ruichuan An and Junkang Zhou and Chang Li and Xiuyuan Lu and Yichi Zhang and Bo Wang and Yuhang Wu and Sen Cui and Miao Liu},
  year={2026},
  eprint={2609.35763},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2609.35763}
}