[
  {
    "path": "README.md",
    "content": "# One Diffusion Step to Real-World Super-Resolution via Flow Trajectory Distillation\n\n[Jianze Li], [Jiezhang Cao](https://www.jiezhangcao.com/), [Yong Guo](https://www.guoyongcs.com/), [Wenbo Li](https://fenglinglwb.github.io/), and [Yulun Zhang*](http://yulunzhang.com/), \"One Diffusion Step to Real-World Super-Resolution via Flow Trajectory Distillation\", ICML, 2025\n\n[[project]] [[arXiv](https://arxiv.org/abs/2502.01993)] [supplementary material] [pretrained models]\n\n\n\n#### 🔥🔥🔥 News\n\n- **2025-02-03:** This repo is released.\n\n---\n\n> **Abstract:** Diffusion models (DMs) have significantly advanced the development of real-world image super-resolution (Real-ISR), but the computational cost of multi-step diffusion models limits their application. One-step diffusion models generate high-quality images in a one sampling step, greatly reducing computational overhead and inference latency. However, most existing one-step diffusion methods are constrained by the performance of the teacher model, where poor teacher performance results in image artifacts. To address this limitation, we propose FluxSR, a novel one-step diffusion Real-ISR technique based on flow matching models. We use the state-of-the-art diffusion model FLUX.1-dev as both the teacher model and the base model. First, we introduce Flow Trajectory Distillation (FTD) to distill a multi-step flow matching model into a one-step Real-ISR. Second, to improve image realism and address high-frequency artifact issues in generated images, we propose TV-LPIPS as a perceptual loss and introduce Attention Diversification Loss (ADL) as a regularization term to reduce token similarity in transformer, thereby eliminating high-frequency artifacts. Comprehensive experiments demonstrate that our method outperforms existing one-step diffusion-based Real-ISR methods.\n\n![](figs/teaser.png)\n\n---\n\n### Pipeline\n\n![](figs/pipeline.png)\n\n---\n\n## 🔖 TODO\n\n- [ ] Release testing code and pre-trained models.\n- [ ] Release training code. \n- [ ] Release pre-trained models.\n- [ ] Provide HuggingFace demo.\n\n## 🔗 Contents\n\n1. Models\n1. Training\n1. Testing\n1. [Results](#results)\n1. [Citation](#citation)\n1. [Acknowledgements](#acknowledgements)\n\n## <a name=\"results\"></a>🔎 Results\n\nWe achieve impressive performance on Real-world Image Super-Resolution. The full results could be downloaded here: [Google Drive](https://drive.google.com/drive/folders/1olqumLOpazfSF4TGTFplO6mF0xKIWdBI?usp=drive_link)\n\n<details>\n<summary>Quantitative Results (click to expand)</summary>\n\n- Results in Table 1 of the main paper\n\n<p align=\"center\">\n  <img width=\"900\" src=\"figs/table1.png\">\n</p>\n\n- Results in Table 2 (RealSet65 testset) of the main paper\n\n<p align=\"center\">\n  <img width=\"450\" src=\"figs/table2.png\">\n</p>\n\n- Quantitative results (×4) on the Real-ISR testset with ground truth.\n\n| Datasets | PSNR ↑ | SSIM ↑ | LPIPS ↓ | DISTS ↓ | MUSIQ ↑ | MANIQA ↑ | TOPIQ ↑ | QAlign ↑ |\n|----------|-------|-------|--------|--------|--------|---------|--------|---------|\n| RealSR   | 24.83  | 0.7175 | 0.3200  | 0.1910  | 68.95   | 0.5335   | 0.6699  | 4.3781   |\n| DRealSR  | 25.92  | 0.7592 | 0.3418  | 0.1628  | 37.82   | 0.5310   | -       | 4.3356   |\n\n- Quantitative results (×4) on the Real-ISR testset without ground truth.\n\n| Datasets  | MUSIQ ↑ | MANIQA ↑ | TOPIQ ↑ | QAlign ↑ |\n|-----------|--------|---------|--------|---------|\n| RealLR200 |  71.60  |  0.5588  | 0.6814  |  4.4004  |\n| RealLQ250 |  72.65  |  0.5490  | 0.6848  |  4.4077  |\n\n\n</details>\n\n<details>\n<summary>Qualitative Results (click to expand)</summary>\n\n- Results in Figure 5 of the main paper\n\n<p align=\"center\">\n  <img width=\"900\" src=\"figs/visual.png\">\n</p>\n\n\n</details>\n\n## <a name=\"citation\"></a>📎 Citation\n\nIf you find the code helpful in your research or work, please cite the following paper(s).\n\n```\n@inproceedings{li2025one,\n  title={One Diffusion Step to Real-World Super-Resolution via Flow Trajectory Distillation},\n  author={Li, Jianze and Cao, Jiezhang and Guo, Yong and Li, Wenbo and Zhang, Yulun},\n  booktitle={ICML},\n  year={2025}\n}\n```\n\n## <a name=\"acknowledgements\"></a>💡 Acknowledgements\n\nThis project is based on [FLUX](https://github.com/black-forest-labs/flux).\n"
  }
]