Open real-world restoration
The official release describes RealRestorer as a real-world image restoration system built on large-scale image editing models, with an emphasis on preserving original scene structure, semantic content, and fine-grained details under practical degradations.
Why RealRestorer
Benchmark gallery
RealRestorer is released with RealIR-Bench and a task prompt set covering common real-world degradations. This section replaces a live demo with a static benchmark-oriented gallery so the page stays aligned with official public assets.
Official references
Blur removal is one of the official restoration targets shown in the released prompt set and benchmark materials.
Low-light enhancement is framed as restoring normal brightness and clarity while retaining the original scene.
Rain removal is presented as a clarity-recovery task in the official example prompts and evaluation flow.
Reflection removal appears in both the task list and the benchmark evaluation examples released with RealRestorer.
Benefits
The official materials frame RealRestorer as an open research release for real degradations, not just a narrow toy benchmark or a closed product demo.
The paper positions RealRestorer as a response to limited training distributions in prior restoration systems, aiming to handle broader real-world degradations more reliably.
The release emphasizes restoration that keeps original scene structure, semantic content, and fine-grained details instead of trading fidelity for aggressive cleanup.
RealRestorer ships with official code, a model card, prompt examples, a degradation pipeline, and the RealIR-Bench benchmark so the evaluation path is visible end to end.
Tasks
The current public release enumerates nine task prompts. This section mirrors those task names rather than inventing extra vertical use cases or unsupported claims.
Features
These are the pieces officially available today across the project page, GitHub repository, model card, and benchmark release.
The release is anchored by the paper “RealRestorer: Towards Generalizable Real-World Image Restoration with Large-Scale Image Editing Models” and its official project page.
The repository publishes the RealRestorer code, evaluation script, example prompts, and instructions for the local patched diffusers checkout.
The model card exposes the official checkpoint, prompt examples, recommended inference settings, and resource links.
RealIR-Bench is the official benchmark release paired with RealRestorer for comparing restoration quality under real degradations.
The published quick start recommends CUDA, torch dtype bfloat16, 28 inference steps, guidance scale 3.0, and seed 42 for the official pipeline.
The repository also includes a degradation pipeline for synthesizing restoration targets such as blur, haze, noise, rain, moire, and reflection.
FAQ
Short answers based on the official paper abstract, repository README, and Hugging Face model card.
RealRestorer is a real-world image restoration model built on top of large-scale image editing models. The official description emphasizes restoring degraded real images while preserving scene structure, semantic content, and fine-grained details.
The current public prompt list covers nine restoration targets: blur, compression artifacts, lens flare, moire, dehazing, low-light enhancement, denoising, rain removal, and reflection removal.
The official code is hosted at GitHub under yfyang007/RealRestorer, and the model checkpoint is published on Hugging Face as RealRestorer/RealRestorer.
RealIR-Bench is the benchmark released alongside RealRestorer for evaluating restoration outputs under real-world degradations. The public dataset page presents it as an image-to-image benchmark tied to the paper.
The official quick start uses the published pipeline with CUDA, torch dtype bfloat16, 28 inference steps, guidance scale 3.0, and seed 42. The repo documents both Diffusers usage and a CLI entrypoint for local inference.
The Hugging Face model card says the code is intended to be released under the Apache License 2.0, while the model and benchmark assets are intended for non-commercial academic research use only. Upstream base-model licenses still apply.