What is AI image restorer suitable for?
This restoration tool uses Transformers.js image-to-image pipelines with Swin2SR real-world and compressed-image models. The real-world model targets blur, noise, and softened detail, while the compressed model is tuned for JPEG compression and web artifacts.
Processing runs in a browser Web Worker with WebGPU fp32 preferred and WASM q8 as the fallback. Model files download on first use and are then reused while the tab stays open. The source image, restoration result, and exported PNG never leave your device.
Supported formats: JPG, PNG, WebP; single file limit: 30 MB; output up to 16 MP and 8192 px per side.
How to use AI image restorer
Step 1
Add a degraded image
Choose a JPG, PNG, or WebP file up to 30 MB.
Step 2
Choose restoration
Select real-world repair or compressed-image recovery based on the damage.
Step 3
Review and download
Restore locally, compare the original with the 4× PNG result, then save it.
FAQ
Does it remove every type of damage?
No. Swin2SR can improve blur, compression artifacts, and low-detail regions, but it cannot reliably recover information that is missing from the source.
Is the image sent to a server?
No. The original image and restored PNG stay in your browser. Only the public ONNX model files are downloaded from Hugging Face.
Why does the result become 4× larger?
Both restoration models reconstruct at 4× scale. The tool checks the projected dimensions first and blocks outputs above 16 megapixels or 8192 pixels per side.
Technology and data sources
- Transformers.js — Browser inference API and WebGPU/WASM runtime documentation.
- Swin2SR Real-World 4× — 4× BSRGAN-PSNR model for blur, noise, and general degradation.
- Swin2SR Compressed 4× — 4× model tuned for heavily compressed images.
- Swin2SR project — Upstream super-resolution architecture and research implementation.
