---
language:
- en
license: cc-by-nc-4.0
task_categories:
- image-to-image
- text-to-image
pretty_name: FreeStyle Bench
tags:
- benchmark
- style-transfer
- image-editing
- cref
- sref
---

# FreeStyle Bench

[Paper](https://huggingface.co/papers/2606.20506) | [Code](https://github.com/Blue2Giant/FreeStyle) | [Project Page](https://blue2giant.github.io/FreeStyle/)

This repository contains benchmark inputs for evaluating **FreeStyle**, a framework for style-content dual-reference generation.

It includes two benchmark subsets:

- `sref_bench/`: a style-transfer benchmark.
- `cref_sref_bench/`: a content-and-style dual-reference benchmark.

Both subsets use the same basic layout: a content-reference folder, a style-reference folder, and a `prompts.json` file that links each pair to its text prompt.

## Repository Layout

```text
<repo-root>/
  README.md
  sref_bench/
    cref/
      ... content reference images ...
    sref/
      ... style reference images ...
    prompts.json
  cref_sref_bench/
    cref/
      ... content reference images ...
    sref/
      ... style reference images ...
    prompts.json
```

## Subset Details

### `sref_bench`

`sref_bench` is the style-transfer benchmark. For each benchmark pair:

- `sref_bench/cref/` contains the content reference image.
- `sref_bench/sref/` contains the style reference image.
- `sref_bench/prompts.json` contains the text prompt or instruction for the pair.

### `cref_sref_bench`

`cref_sref_bench` is the content-and-style dual-reference benchmark. It has the same structure as `sref_bench`, where each key in `prompts.json` is a pair basename used to align the content reference image with the style reference image.

## Sample Usage

You can use the following code to read a benchmark pair and its associated metadata:

```python
import json
from pathlib import Path
from PIL import Image

bench_dir = Path("/path/to/FreeStyle_Bench/sref_bench")  # or cref_sref_bench

with open(bench_dir / "prompts.json", "r", encoding="utf-8") as f:
    prompts = json.load(f)

pair_name, prompt = next(iter(prompts.items()))

# Find images by matching the pair basename. The actual extension may vary.
cref_path = next((bench_dir / "cref").glob(pair_name + ".*"))
sref_path = next((bench_dir / "sref").glob(pair_name + ".*"))

content_ref = Image.open(cref_path).convert("RGB")
style_ref = Image.open(sref_path).convert("RGB")

print("pair:", pair_name)
print("prompt:", prompt)
print("content reference:", cref_path)
print("style reference:", sref_path)
```

## Notes

- `cref` means content reference.
- `sref` means style reference.
- `prompts.json` is the index file for each subset.
- The key of each `prompts.json` entry is the pair basename used to match images in `cref/` and `sref/`.

## Citation

```bibtex
@article{lan2026freestyle,
  title         = {FreeStyle: Free Control of Style-Content Dual-Reference Generation from Community LoRA Mining},
  author        = {Lan, Jinghong and Cheng, Wei and Chen, Yunuo and Ye, Ziqi and Xing, Peng and Fang, Yixiao and Wang, Rui and Yang, Yufeng and Zhang, Xuanyang and Zou, Difan and Zeng, Xianfang and Yu, Gang and Zhang, Chi},
  journal       = {arXiv preprint arXiv:2606.20506},
  year          = {2026}
}
```