Blurry Photo Savior: Devours 9 Types of Degradation in One Go, Open-Source Leader Closing In on Closed-Source Ceiling

You took a night shot. Hand shook. It's blurry.

You think: forget it, can't fix it anyway. Deblurring tools only deblur, denoising can't handle raindrops, rain removal is useless in low light. Nine diseases require nine specialists; by the time you finish registration, the photo is already cold.

Then you open RealRestorer, type "Please deblur the image and make it sharper". 28 inference steps, 3 seconds output. Clear, clean, structure unchanged.

Deblurring, denoising, deraining, deglaring, dehazing, decompression artifact removal, demoiréing, reflection removal, low-light enhancement — one model handles it all. Not nine models stitched together, but one model that swallows nine degradations.

What Image Restoration Was Like Before

Like a doctor who only knows how to treat a cold. You have a fever, he gives you antipyretics. You also cough, have a runny nose, and muscle aches? He's lost.

Traditional image restoration models are such "specialists." Train one for deblurring, another for denoising, another for deraining. Models are irrelevant to each other. When faced with mixed degradation — like a night view that is blurry, dark, and noisy — you either feed them one by one or just give up.

The problem with feeding one by one is that the output after the first model has already deviated from the original distribution. The second model receives not a "truly degraded image" but the "residual of the first model's restoration." Errors accumulate layer by layer; by the third round, details are lost.

RealRestorer's approach is completely different. It does not treat "a certain degradation"; it treats "degradation" itself.

How It Works

Built on a large-scale image editing model. Not training a small network from scratch to learn deblurring, but performing restoration on top of a large model that has already seen massive amounts of images and understands scene semantics.

What does this mean? It means it knows what the photo should look like.

A small deblurring model doesn't know how much texture the leaves in the photo should have or what skin texture should be like — it only knows how to sharpen edges. But a large model that understands scene semantics can restore reasonable details while deblurring, because it has "seen" what clear leaves and skin look like.

The training uses a two-stage approach. The first stage learns degradation removal capability; the second stage fine-tunes stability with real degradation data. The key is the Progressively-Mixed strategy — the second stage isn't purely real data; it retains a small amount of synthetic degradation data mixed in. Prevents overfitting. Pure real data tends to drift, pure synthetic data isn't realistic enough; mixing works just right.

Prompt-driven operation is another highlight. Instead of a fixed pipeline, natural language controls the restoration behavior:

Same image, different prompt, different restoration direction. Previously, this required nine models; now just one sentence.

Benchmark: Open-Source First, Global Third

RealIR-Bench is released together by the RealRestorer team, covering 9 types of real degradation. On this benchmark:

• Ranked first among open-source methods
• Ranked third globally, behind Nano Banana Pro and GPT-Image-1.5
• Best performance in deblurring and low-light enhancement — two notoriously hardest tasks

"Open-source first" is not surprising. What's surprising is that the gap with top closed-source models is almost negligible.

In the past, there was a huge gap between open-source and closed-source models. Use open-source, cheap but worse results; use closed-source, better results but uncontrollable, pay, and your data must be uploaded. RealRestorer, for the first time, narrows this gap to the point of "do you really care about that 1%?".

Hands-On: Just Run It

Code, models, and benchmark are fully open-sourced on GitHub. Inference is simple:

python3 infer_realrestorer.py \
  --model_path /path/to/realrestorer_bundle \
  --image /path/to/input.png \
  --prompt "Restore the details and keep the original composition." \
  --output /path/to/output.png \
  --device cuda \
  --torch_dtype bfloat16 \
  --num_inference_steps 28 \
  --guidance_scale 3.0

28 inference steps, guidance_scale 3.0, bfloat16 precision. Not slow, GPU-friendly.

Apple Silicon MPS inference is supported — Mac users can also run. Diffusers inference and ComfyUI nodes are community-supported.

The accompanying degradation_pipeline and RealIR-Bench are also open-sourced. Want to run your own comparison experiments or train your own restoration model? The evaluation pipeline is end-to-end visible.

StepFun Strikes Again

The author team is from Southern University of Science and Technology, StepFun, and Shenzhen Institute of Advanced Technology, CAS.

The name StepFun has been appearing frequently lately. The core author of Baidu Unlimited OCR, Wei Haoran, also comes from StepFun. The company's accumulation in visual understanding is deeper than the outside perceives.

The first author of RealRestorer is Yufeng Yang. The GitHub repository has 291 Stars, 25 Commits, 8 Open Issues — the community is using it, and feedback is being addressed. It's not one of those projects that publishes a paper and runs away.

One Model vs. Nine Models: A Strategic Divide

On the surface, RealRestorer is an image restoration tool. But looking deeper, it represents a strategic choice: general vs. specialized.

Specialized models can achieve the best on a single task, but struggle with mixed real-world degradation. General models sacrifice a bit of single-task performance to cover real-world scenarios.

The real demand for image restoration is never a clean category like "deblurring" or "denoising." The bad photos you take with your phone often have blur, noise, low light, and compression artifacts simultaneously. Users want one button, not nine.

RealRestorer proves that the general route based on large-scale image editing models can already approach the closed-source ceiling in the open-source world. The next step is to see if it can eat away that last 1% gap — or if closed-source models will pull away again.

You took another night shot. Hand still shook.

But this time, you didn't panic.

GitHub: https://github.com/yfyang007/RealRestorer

HuggingFace: https://huggingface.co/RealRestorer/RealRestorer

Website: https://realrestorer.com