Restore images

Which models are the fastest for restoring images?

If you need quick results, tencentarc/gfpgan and cjwbw/bigcolor are among the fastest options in the image-restoration collection. They can process typical images in just a few seconds, which makes them great for quick cleanup of old or low-quality photos.
Faster models can sometimes leave minor artifacts, especially on heavily damaged images.

Which models offer the best balance of quality and flexibility?

jingyunliang/swinir is a strong all-around model for general image restoration — it handles denoising, deblurring, and improving low-resolution images with consistent results.
For more specialised tasks, models like piddnad/ddcolor can add color to black-and-white photos, while megvii-research/nafnet is good for restoring sharpness and structure without over-smoothing.

What works best for restoring old or damaged photos with faces?

If your image includes faces — like family portraits or old photographs — tencentarc/gfpgan is designed for this exact purpose. It reconstructs facial features realistically even when the original image is blurry or heavily degraded.
You can pair it with another restoration model like jingyunliang/swinir if you want to clean up the background or sharpen other areas.

What’s best if I want to add color to vintage black-and-white images?

For faithful colorisation of grayscale images, piddnad/ddcolor is the top choice in the image-restoration collection. It can handle portraits, landscapes, and even artistic stills with good accuracy.
If you prefer more varied or stylised color options, cjwbw/bigcolor generates multiple alternative colorisations from the same input.

How do the main types of image-restoration models differ?

• Face restoration models like tencentarc/gfpgan are focused on reconstructing facial details while keeping them natural.
• General restoration models such as jingyunliang/swinir and megvii-research/nafnet clean up noise, fix blur, and enhance structure across the whole image.
• Colorisation models like piddnad/ddcolor or cjwbw/bigcolor add or enhance color information after restoration.
• Some models are speed-oriented, while others focus on fine detail and quality.

What kinds of outputs can I expect?

Most models in this collection return restored image files in PNG or JPG format. Depending on the model, you might see:

• Sharper, clearer faces.
• Reduced noise and blur.
• Improved resolution and structure.
• Realistic or stylised color added to grayscale images.
  Each model focuses on different aspects of restoration, so results will vary.

How can I self-host or publish my own image-restoration model?

You can package your own restoration model with Cog and push it to Replicate. This lets you set up inputs, outputs, and configuration while deciding whether to keep it private or share it with others.

Can I use image-restoration models for commercial work?

Many models in the image-restoration collection allow commercial use, but license terms vary. Always check the model’s license and attribution requirements before using outputs in commercial projects.

How do I use these models on Replicate?

1. Choose a model from the image-restoration collection.
2. Upload your image or provide a URL.
3. Select options like face restoration, colorisation, or enhancement if available.
4. Run the model and download the restored output.
5. Use the restored image in your workflow — for printing, archiving, or creative projects.

What should I know before restoring an image?

• The better your input scan or photo, the better the result.
• Face restoration works best when faces are at least partially visible.
• Colorisation works best when the base image is clean — consider running a restoration model first.
• Larger images may take longer to process.
• Restoration can recover or enhance details, but it won’t perfectly recreate lost information.
• Always inspect outputs carefully, especially for historical or sentimental photos.