Recent advances in diffusion-based real-world image super-resolution (Real-ISR) have demonstrated remarkable perceptual quality, yet the balance between fidelity and controllability remains a problem: multi-step diffusion-based methods suffer from generative diversity and randomness, resulting in low fidelity, while one-step methods lose control flexibility due to fidelity-specific finetuning.
ODTSR addresses this by presenting a one-step diffusion transformer based on Qwen-Image that performs Real-ISR considering fidelity and controllability simultaneously. It introduces a newly designed Noise-hybrid Visual Stream (NVS) that receives low-quality images with adjustable noise (Control Noise) and consistent noise (Prior Noise). Furthermore, Fidelity-aware Adversarial Training (FAA) is employed to enhance controllability and achieve one-step inference. ODTSR not only achieves state-of-the-art (SOTA) performance on generic Real-ISR, but also enables prompt controllability on challenging scenarios such as real-world scene text image super-resolution (STISR) of Chinese characters without training on specific datasets.
Key Features
One-Step Super-Resolution: Based on Qwen-Image, ODTSR trains a single-step SR model using LoRA, with model parameters reaching 20B.
Controllability: With our proposed Noise-hybrid Visual Stream and Fidelity-aware Adversarial Training, the SR process can be jointly controlled by prompts as well as a Fidelity Weight $f$.
Multilingual Support: English and Chinese prompts are supported.
Versatile Performance: The model demonstrates strong performance in text images, fine-grained textures, and face images.
Visual Results
Results with fixed prompts & high fidelity
Results with fixed prompts & high fidelity
Under the high-fidelity setting with a fixed prompt, our model produces restorations that adhere more closely to the LQ input while remaining natural, significantly reducing the sense of AI processing.
Text Real-ISR Results
Text Real-ISR Results
In text scenarios, when the prompt specifies the text to be restored, the model automatically matches the LQ text and performs the restoration.
Controllable Real-ISR Results
Controllable Real-ISR Results
Qualitative results of controllable SR with prompt and adjustable Fidelity Weight (denoted as $f$) on Div2k-val dataset. As $f$ decreases from 1 to 0, detail generation and prompt adherence gradually strengthen.
(For training) Download base model to your disk: Wan2.1-T2V-1.3B
(For inference) Download the trained ODTSR model weight: huggingface
Inference with Script
Note: you need at least 40GB GPU memory to infer. We will support CPU offload to reduce GPU memory usage soon.
We now supports tile-based processing (tile size: 512×512), enabling input of arbitrary resolutions and SR at any scale factor.
Please replace experiments/qwen_one_step_gan/${EXP_DATE}/checkpoints/net_gen_iter_10001.pth with the trained ODTSR model weight.
This project is based on DiffSynth-Studio.
We also leveraged some of PiSA-SR's code in dataloader part.
Thanks for the awesome work!
Citation
If ODTSR is helpful to you, please consider citing our paper:
bibtex
1@article{fang2025onestep,
2 title={One-Step Diffusion Transformer for Controllable Real-World Image Super-Resolution},
3 author={Fang, Yushun and Chen, Yuxiang and Yin, Shibo and Hu, Qiang and Yao, Jiangchao and Zhang, Ya and Zhang, Xiaoyun and Wang, Yanfeng},
4 journal={arXiv preprint arXiv:2511.17138},
5 year={2025}
6}
