ASASR · Faithful x4 Image Super-Resolution
Coloring the Noise: Adversarial Sobolev Alignment for Faithful Image Super-Resolution (ICML 2026)
ASASR turns a low-resolution image into a faithful 512 × 512 reconstruction using a FLUX.1-dev backbone with dual-LoRA inference — a base SR LoRA plus an AS-DPO alignment LoRA. Upload a roughly 128 × 128 image (or any image — it will be center-cropped and downsampled), then run the sampler and compare the input and output side by side.
Example inputs (click to run)
License: CC-BY-NC-4.0 — non-commercial research use only. ASASR weights additionally inherit the non-commercial terms of FLUX.1-dev.
@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}
}