Xixixixihahahaha/RealAlign-Dataset

RealAlign Preference Dataset

This is the curated preference dataset for the paper "When Preference Labels Fall Short: Aligning Diffusion Models from Real Data" (ICML 2026).

Resource	Link
📄 Paper	arXiv:2605.19839
🌐 Project page	cwyxx.github.io/RealAlign
🚀 Code	github.com/Cwyxx/RealAlign

Dataset summary

RealAlign studies whether real data can replace manually annotated preference labels for aligning text-to-image diffusion models. Instead of comparing two model-generated images and labelling which is preferred, we treat a high-quality reference image as the preferred ("win") sample — a real photograph (HPDv3) or a curated high-quality generated image (Civitai-top) — and contrast it with a perturbed / inpainted version as the non-preferred ("lose") sample. This yields preference pairs without any human preference annotation, and we show such reference-based supervision aligns diffusion models comparably to existing preference-based methods.

This dataset provides the resulting (real, fake) preference pairs used to train the RealAlign SD-1.5 and SD-3.5-M models.

Sources

This release contains two sources, kept as separate subsets:

Subset	Reference (win) images	Prompt / image origin
HPDv3/	real photographs	Human Preference Dataset v3
Civitai-top/	high-quality generated images	Civitai top SFW images

Note: The Pick-a-Pic v2 subset used in the paper is not included in this release because the source data may contain not-safe-for-work (NSFW) content.

Curation pipeline

The pairs are produced by RealAlign's four-stage data-curation pipeline (see data_curation/ in the code repo):

1. Extract — collect (uid, prompt) entries for each source.
2. Construct pairs — for each reference image, generate the non-preferred ("fake") counterpart by computing a U²-Net saliency mask and inpainting the salient region with a text-to-image model (SD / SD-3.5 / PixArt). The real image becomes the "win" sample and the inpainted image the "lose" sample for the same prompt.
3. Score — score candidates with colorfulness, PickScore, and a Qwen3-VL anime classifier.
4. Filter — curate per source. HPDv3 uses anime drop → color filter → discard negative → top-512; Civitai-top applies top selection only.

Directory layout

HPDv3/
├── HPDv3.csv          # uid, prompt, win_image_path, lose_image_path
├── real/<uid>.png     # reference (preferred / "win") images
└── fake/<uid>.png     # perturbed (non-preferred / "lose") images
Civitai-top/
├── Civitai-top.csv
├── real/<uid>.png
└── fake/<uid>.png

Each subset ships 512 final curated preference pairs as the images under real/ and fake/, paired by uid. The accompanying CSV may list additional candidate rows from earlier curation stages (HPDv3 in particular), so treat the images present in real//fake/ as the delivered set and intersect the CSV by uid with those files.

CSV columns

Column	Description
uid	Pair identifier; matches the <uid>.png filenames in real/ and fake/.
prompt	Text prompt for the pair.
win_image_path	Path to the preferred (real) image.
lose_image_path	Path to the non-preferred (fake / inpainted) image.

Note: win_image_path / lose_image_path in the CSVs are the absolute paths from the original training machine (e.g. /data_center/.../real/<uid>.png). When using this dataset, resolve images by uid against the local real/ and fake/ folders, or remap the path prefix to your download location.

Usage

The curated CSV is consumed directly by the RealAlign trainers (csv_file_path_train for SD-1.5, config.{irl,dpo}.csv_file_path for SD-3.5-M). See the GitHub repository for full training instructions.

Citation

If you find this dataset useful, please consider citing:

@article{chen2026preference,
  title={When Preference Labels Fall Short: Aligning Diffusion Models from Real Data},
  author={Chen, Weiyan and Deng, Weijian and Xiao, Yao and Tu, Weijie and Dong, ZiYi and Radwan, Ibrahim and Lin, Liang and Wei, Pengxu},
  journal={arXiv preprint arXiv:2605.19839},
  year={2026}
}