PixRestore - A Unified Image Restoration Model Jointly Launched by The Hong Kong Polytechnic University and OPPO

What is PixRestore

PixRestore is a unified image restoration model jointly launched by The Hong Kong Polytechnic University and OPPO Research Institute. The model uses a pixel-space Diffusion Transformer to perform flow matching directly at the pixel level, avoiding detail loss. Through DINO-driven adaptive layer feature routing, the lightweight model with only 53.7M parameters can efficiently handle 8 degradation tasks such as super-resolution, denoising, deblurring, deraining, and dehazing, with single-step inference taking only 44 milliseconds, achieving leading levels in both fidelity and efficiency.

Main Features of PixRestore

Super-Resolution: Restore low-resolution images to high-definition images with rich details.
Image Denoising: Remove random noise particles in the image and restore clean textures.
Motion Deblurring: Eliminate blur caused by motion or defocus and restore sharp edges.
Rain Streak Removal: Remove the interference of dense rain streaks on the image and restore the background content.
Raindrop Removal: Eliminate raindrop occlusion on glass or lens surfaces and restore a clear view.
Image Dehazing: Disperse the haziness caused by atmospheric haze and improve image clarity and color.
Desnowing: Remove snowflake noise in the image and restore the original appearance of the scene.
Low-Light Enhancement: Improve brightness and details in low-light environments and enhance image visibility.

Technical Principles of PixRestore

Pixel-Space Diffusion Transformer: PixRestore performs flow matching directly at the pixel level, completely abandoning the VAE encoder to avoid the loss of high-frequency details caused by latent space compression, while also eliminating dependence on T2I pretrained models, ensuring that restoration results are highly consistent with the original content.
Adaptive Layer Feature Routing: The model uses the DINO feature similarity of LQ-HQ images to dynamically evaluate the reliability of features from each layer of the visual encoder; reliable layers are fused into dense conditions and injected into the DiT, while stronger high-quality feature supervision is applied to unreliable layers, thereby adaptively removing different degradations.
Single-Step Efficient Inference: After full training, PixRestore is further fine-tuned into a single-step generator through a DINO-based adversarial objective. With only 53.7M parameters, it can complete inference in 44ms, achieving the best balance among fidelity, perceptual quality, and computational efficiency.

Follow WeChat and reply "Open Source" to joinAI Open Source Project Communication Group

How to Use PixRestore

Environment Setup: Clone the PixRestore code repository from GitHub and install the required dependencies.
Model Acquisition: Download the pretrained model weight files from Hugging Face.
Image Preparation: Prepare the low-quality image to be restored, supporting 8 types of degradation such as super-resolution, denoising, and deblurring.
Run Inference: Run the inference script. The model performs single-step inference by default and can output the restoration result in about 44ms.
Save Results: Obtain and save the high-quality restored image for direct use in subsequent applications.

Core Advantages of PixRestore

Pixel-Level Fidelity: Diffusion modeling is performed directly in pixel space, completely avoiding the loss of high-frequency details and content artifacts caused by VAE encoding.
Extremely Lightweight and Efficient: With only 53.7M parameters, single-step inference takes only 44 milliseconds, and the computational overhead is far lower than mainstream latent diffusion restoration solutions.
Unified Multi-Task Coverage: A single model can handle 8 types of degradation, including super-resolution, denoising, deblurring, deraining, dehazing, desnowing, and low-light enhancement.
Adaptive Degradation Perception: Dynamically evaluates the reliability of each layer through DINO feature similarity, achieving an intelligent trade-off between degradation removal and detail preservation.
Independent Training Paradigm: It does not rely on any text-to-image pretrained model and is trained completely from scratch, avoiding content bias introduced by generative priors.
Comprehensive Performance Leadership: On public benchmarks and real-world test sets, fidelity, perceptual quality, and degradation removal capability all outperform existing unified restoration models.

PixRestore Project Addresses

Project Website：https://csslc.github.io/pixrestore-page
GitHub Repository：https://github.com/csslc/PixRestore
HuggingFace Model Library：https://huggingface.co/VCLab-PolyU/PixRestore
arXiv Technical Paper：https://arxiv.org/pdf/2608.16793

Comparison of PixRestore with Similar Competing Products

Comparison Dimension	PixRestore	PixelHacker
Core Task	Unified image restoration (UIR): removes 8 types of degradation such as super-resolution, denoising, deblurring, deraining, dehazing, desnowing, and low-light enhancement	Image inpainting: fills missing regions in an image according to a mask, restoring structural and semantic consistency
Technical Architecture	Pixel-space Diffusion Transformer, no VAE, trained from scratch	Latent Diffusion Model, based on a VAE encoder, converting features from pixel space to latent space
Parameter Count	53.7M(lightweight)	Specific parameters are not disclosed; based on a standard diffusion U-Net, the parameter count is usually in the hundreds of millions
Inference Efficiency	Single step 44ms, with a computational cost of about 658G FLOPs	Multi-step iterative denoising requires tens of sampling steps, resulting in significantly higher inference latency.
Conditional input	Only low-quality (LQ) images are required, with automatic perception of degradation types.	Requires an input image + mask to explicitly specify the regions to be restored.
Training data	Large-scale paired data of diverse scenes and degradations.	14 million image-mask pairs, based on foreground (116 classes) and background (21 classes) annotations.
Feature injection method	DINO-driven adaptive layer feature routing, dynamically fusing multi-layer visual features.	Latent class guidance (LCG), intermittently injecting foreground/background embeddings through linear attention.

Application scenarios of PixRestore

Real-time photo enhancement on mobile phones: With an ultra-low latency of 44ms, it restores night scene noise, motion blur, and raindrop occlusion in real time on mobile devices such as OPPO, improving the quality of the final images.
Autonomous driving and drone vision: In harsh weather such as rain, fog, and snow, it uniformly removes image degradations, restores details of road signs and obstacles, and enhances the reliability of perception systems.
Surveillance video quality restoration: Performs super-resolution and denoising on old or low-bitrate surveillance footage, restoring key information such as faces and license plates to assist security evidence collection.
Digitization of old photos and historical archives: Uniformly repairs fading, scratches, mildew spots, and low-light issues, efficiently converting precious historical images into high-fidelity digital archives.
Medical imaging enhancement: Improves the imaging quality of endoscopes, microscopes, and other devices under low-light conditions, removes noise and blur, and helps doctors observe lesion details more clearly.