README.md

---
license: cc-by-nc-4.0
tags:
  - super-resolution
  - image-super-resolution
  - flux
  - lora
  - dpo
  - diffusion
library_name: diffusers
pipeline_tag: image-to-image
---

# ASASR — Coloring the Noise: Adversarial Sobolev Alignment for Faithful Image Super-Resolution

Pretrained weights for the ICML 2026 paper
**[Coloring the Noise: Adversarial Sobolev Alignment for Faithful Image Super-Resolution](https://arxiv.org/abs/2605.23264)**
(Hongbo Wang, Huaibo Huang, Pin Wang, Jinhua Hao, Chao Zhou, Ran He).

➡️ **Code & full instructions: https://github.com/wafer-bob/ASASR**

ASASR performs ×4 image super-resolution with a **FLUX.1-dev** backbone and dual-LoRA
inference: a base **SR LoRA** (upscaling prior, OminiControl) plus our **DPO LoRA** trained
with a Sobolev frequency-weighted, adversarially-guided DPO objective (AS-DPO).

## Files

| File | Size | Use |
|---|---|---|
| `sr_lora/pytorch_lora_weights_v2.safetensors` | ~885 MB | base SR LoRA — **inference** |
| `dpo_lora/adapter_model.safetensors` | ~111 MB | ASASR AS-DPO LoRA — **inference** |
| `adv_lora/adapter_model.safetensors` | ~111 MB | rank-16 AMG adversary — **training only** |

## Download

```bash
huggingface-cli download wafer-bob/ASASR --local-dir ./checkpoints
```

Then follow the [GitHub README](https://github.com/wafer-bob/ASASR) for inference
(`bash scripts/infer.sh`) and training.

## License

This project is released under [CC-BY-NC-4.0](https://creativecommons.org/licenses/by-nc/4.0/) for **non-commercial research use only**.

Copyright (c) 2026 The Authors and Kuaishou Technology.

## Citation

```bibtex
@inproceedings{wang2026asasr,
  title     = {Coloring the Noise: Adversarial Sobolev Alignment for Faithful Image Super-Resolution},
  author    = {Wang, Hongbo and Huang, Huaibo and Wang, Pin and Hao, Jinhua and Zhou, Chao and He, Ran},
  booktitle = {International Conference on Machine Learning (ICML)},
  year      = {2026}
}
```