README.md

---
license: mit
base_model: stabilityai/stable-diffusion-3.5-medium
library_name: peft
tags:
  - semantic-segmentation
  - dataset-distillation
  - text-to-image
  - lora
  - s2rl
  - arxiv:2610.35491
---

<p align="center">
  <img src="https://raw.githubusercontent.com/00why00/S2RL/main/assets/banner.svg" alt="S²-RL" width="100%">
</p>

<h1 align="center">⚗️ S²-RL</h1>

<p align="center">
  <strong>Sample-Set Dual Reinforcement Learning<br>for Generative Semantic Segmentation Dataset Distillation</strong>
</p>

<p align="center">
  <a href="https://haoyuwang.com">Haoyu Wang</a> · <a href="https://scholar.google.com/citations?hl=en&amp;user=Sd2wRM4AAAAJ">Fei Zhou</a> · <a href="https://openreview.net/profile?id=~Qingqing_Qiu1">Qingqing Qiu</a> · <a href="https://teacher.nwpu.edu.cn/nwpuzhanglei.html">Lei Zhang</a> · <a href="https://teacher.nwpu.edu.cn/weiwei.html">Wei Wei</a> · <a href="https://gr.xupt.edu.cn/info/1213/7446.htm">Chen Ding</a>
</p>

<p align="center">
  <a href="https://arxiv.org/abs/2610.35491">📄 arXiv</a> ·
  <a href="https://00why00.github.io/S2RL/">🌐 Project Page</a> ·
  <a href="https://github.com/00why00/S2RL">💻 Code</a> ·
  <a href="#datasets">📂 Datasets</a> ·
  <a href="#training">🚀 Policy Training</a> ·
  <a href="#generation">🎨 Generation</a>
</p>

This repository contains the **ADE20K and COCO datasets and LoRA adapters** used in the S²-RL main experiments: 10 data archives, 21,700 image–mask pairs, and two dataset-specific adapters.

<a id="datasets"></a>

## 📂 Datasets

### Selected real-data subsets

The selection baselines provide subsets of real images and their ground-truth masks. Uniform subsets also supply the category prompts for S²-RL.

| Method | ADE20K | COCO |
| :--- | :---: | :---: |
| Random | [Download](https://huggingface.co/0why0/S2RL/resolve/main/datasets/ADE20K/Random.tar.gz?download=true) | [Download](https://huggingface.co/0why0/S2RL/resolve/main/datasets/COCO/Random.tar.gz?download=true) |
| Uniform | [Download](https://huggingface.co/0why0/S2RL/resolve/main/datasets/ADE20K/Uniform.tar.gz?download=true) | [Download](https://huggingface.co/0why0/S2RL/resolve/main/datasets/COCO/Uniform.tar.gz?download=true) |
| Herding | [Download](https://huggingface.co/0why0/S2RL/resolve/main/datasets/ADE20K/Herding.tar.gz?download=true) | [Download](https://huggingface.co/0why0/S2RL/resolve/main/datasets/COCO/Herding.tar.gz?download=true) |
| K-Center | [Download](https://huggingface.co/0why0/S2RL/resolve/main/datasets/ADE20K/K-Center.tar.gz?download=true) | [Download](https://huggingface.co/0why0/S2RL/resolve/main/datasets/COCO/K-Center.tar.gz?download=true) |

All data archives include **three budgets and five sets per budget**, with paired JPEG images and PNG segmentation masks.

| Dataset | Ratios | Image–mask pairs per set |
| :--- | :--- | :--- |
| ADE20K | `0.0025`, `0.005`, `0.01` | 50, 101, 202 |
| COCO | `0.000625`, `0.00125`, `0.0025` | 73, 147, 295 |

Extract archives into the data root used by the [code](https://github.com/00why00/S2RL#data):

```bash
tar -xzf <archive>.tar.gz -C "$S2RL_DATA_ROOT"
```

```text
ILDistillation/<dataset>/<method>/
├── images/time_<n>_ratio_<r>/*.jpg
├── annotations/time_<n>_ratio_<r>/*.png
└── _manifest.json
```

ADE20K masks use labels `1..150`, with `0` ignored. COCO masks use `0..132`, with `255` ignored; class order follows [dataset/coco.py](https://github.com/00why00/S2RL/blob/main/dataset/coco.py).

<a id="training"></a>

## 🚀 Policy Training

Policy training and generation are described in the [code repository](https://github.com/00why00/S2RL#training).

| Dataset | Policy weights | Adapter config |
| :--- | :---: | :---: |
| ADE20K | [Download](https://huggingface.co/0why0/S2RL/resolve/main/checkpoints/ADE20K/adapter_model.safetensors?download=true) | [Download](https://huggingface.co/0why0/S2RL/resolve/main/checkpoints/ADE20K/adapter_config.json?download=true) |
| COCO | [Download](https://huggingface.co/0why0/S2RL/resolve/main/checkpoints/COCO/adapter_model.safetensors?download=true) | [Download](https://huggingface.co/0why0/S2RL/resolve/main/checkpoints/COCO/adapter_config.json?download=true) |

These LoRA adapters use [Stable Diffusion 3.5 Medium](https://huggingface.co/stabilityai/stable-diffusion-3.5-medium) as the base model. Save both files in the same directory and set `POLICY_DIR` in the generation script to that directory.

<a id="generation"></a>

## 🎨 Generation and selection

### Distilled datasets

S²-RL provides distilled datasets of generated images and pseudo-labels for downstream segmentation training.

| Method | ADE20K | COCO |
| :--- | :---: | :---: |
| **S²-RL** | [Download](https://huggingface.co/0why0/S2RL/resolve/main/datasets/ADE20K/S2RL.tar.gz?download=true) | [Download](https://huggingface.co/0why0/S2RL/resolve/main/datasets/COCO/S2RL.tar.gz?download=true) |

Extract into `$S2RL_DATA_ROOT`. Image–mask pairs are stored under `ILDistillation/<dataset>/S2RL/`; use `DATASET_NAME="S2RL"` in the downstream evaluation scripts.

## 🖊️ Citation

```bibtex
@misc{wang2026s2rl,
  title={S$^2$-RL: Sample-Set Dual Reinforcement Learning for Generative Semantic Segmentation Dataset Distillation},
  author={Haoyu Wang and Fei Zhou and Qingqing Qiu and Lei Zhang and Wei Wei and Chen Ding},
  year={2026},
  eprint={2610.35491},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2610.35491}
}
```

## 📄 License

S²-RL contributions are licensed under the [MIT License](https://github.com/00why00/S2RL/blob/main/LICENSE). Third-party components and assets follow their original licenses.