FluxSR - An image super-resolution model jointly developed by Shanghai Jiao Tong University, Huawei, and other universities.
FluxSR is a novel single-step diffusion model developed by Shanghai Jiao Tong University, Harvard University, South China University of Technology, and Huawei Noah's Ark Lab, specifically for real-world image super-resolution (Real-ISR) tasks. FluxSR is based on...
What is FluxSR?
FluxSR is a novel single-step diffusion model developed by Shanghai Jiao Tong University, Harvard University, South China University of Technology, and Huawei Noah's Ark Lab specifically for real-world image super-resolution (Real-ISR) tasks. Based on the FLUX.1-dev text-to-image (T2I) diffusion model, FluxSR uses Flow Trajectory Distillation (FTD) technology to distill the multi-step flow matching model into a single-step super-resolution model. FluxSR's core advantage lies in its ability to efficiently generate high-quality super-resolution images while maintaining the high realism of the T2I model. FluxSR uses TV-LPIPS perceptual loss and Attention Diversification Loss (ADL) to optimize high-frequency details and reduce artifacts. FluxSR demonstrates superior performance on multiple datasets, particularly excelling in no-reference image quality assessment metrics, significantly reducing computational costs and providing a new solution for efficient, high-quality image super-resolution.
Main functions of FluxSR
- High-efficiency single-step super-resolution reconstructionIt efficiently restores low-resolution images to high-resolution images during single-step diffusion, significantly reducing computational costs and inference latency, making it suitable for fast image processing needs.
- High-fidelity image generationExtracting high-fidelity details from a pre-trained text-to-image (T2I) model and applying it to super-resolution tasks generates images with rich detail and high fidelity.
- High-frequency detail recovery and artifact suppressionIt can effectively restore high-frequency details of images and reduce high-frequency artifacts and repetitive patterns.
FluxSR's technical principles
- Flow trajectory distillation(Flow Trajectory Distillation, FTD):
- Noise-to-image stream generationGenerate a stream of noise to an image using a pre-trained T2I model.
- Derivation of low-resolution to high-resolution streamingThe flow trajectory from LR to HR is derived based on mathematical relationships, avoiding the distribution shift caused by directly optimizing the SR flow.
- Training strategy for single-step diffusion modelBased on a training strategy that is friendly to large models, it generates noise-to-image streaming data pairs offline, avoiding reliance on additional teacher models during training. This significantly reduces memory consumption and training costs, making single-step model training more efficient.
- Perceived loss and regularization:
- TV-LPIPS perceived lossCombining total variation (TV) and LPIPS (Learned Perceptual Image Patch Similarity) loss, it emphasizes the recovery of high-frequency components and reduces artifacts in the generated image.
- Loss of attention diversification (ADL)This approach reduces the similarity between different tokens in the transformer, enhances the diversity of attention, and eliminates high-frequency artifacts.
- Efficient ReasoningIt relies on only a single flow model during the inference phase, avoiding the high computational overhead of multi-step diffusion models. Based on FTD technology, it generates high-quality super-resolution images in a single step while maintaining a level of realism comparable to multi-step models.
FluxSR project address
- GitHub repository:https://github.com/JianzeLi-114/FluxSR
- arXiv technical paper:https://arxiv.org/pdf/2502.01993
Application scenarios of FluxSR
- Restoration of old photosIt restores old photos that are low-resolution, blurry, or damaged to high-resolution, clear images.
- Film and television productionIn film and television post-production, low-resolution footage is upscaled to high resolution to meet the demands of high-definition or 4K production.
- Enhanced medical imagesIt improves the resolution of low-resolution medical images (such as X-rays, CT scans, and MRIs) to help doctors diagnose diseases more accurately.
- Smartphone photographyImprove the clarity of low-resolution photos taken with your phone, especially in low-light or fast-moving scenes.
- Quality InspectionIn industrial production, improving the resolution of image inspection systems on production lines helps to detect product defects more accurately.