MFIR - Multi-Frame Image Restoration

Upload multiple degraded frames of the same scene, and the model will fuse them into a single high-quality image.

How it works: The model aligns all frames to a reference frame using deformable convolutions, then fuses them using temporal attention to extract the best information from each frame.

Upload Frames (2-16 images)

Processing Size

Larger = better quality, slower

Reference Frame

Which frame to use as reference

0 15

Restored Output

Status

Examples

Click to load sample images

Tips:

Upload 2-16 frames of the same scene
Frames can have different degradations (blur, noise, etc.)
The model works best when frames have slight variations
Processing size affects quality and speed

Author: Veli Karaarslan | GitHub