Deblur Image (NAFNet)

@megviiresearch1 VersionEnhance

NAFNet restorer that cleans an Image with a task-specific checkpoint (motion deblur, general frame restoration, or real-camera denoise) and emits a restored RGB Image at the same resolution. Use when blur or noise hurts downstream accuracy.

How it fits

Typical backends

Pre-OCR cleanup: a degraded document frame is restored so character recognition runs on a sharp, low-noise image.

An HTTP restoration service: noisy or blurred frames are cleaned and returned over HTTP for display.

Caveats

Accuracymodel selects the training target — one variant for motion deblur, one for general frame restoration, one for real-camera denoise — and applying the wrong variant produces worse output than no restoration, with no runtime warning.
Parameter interactionVery small tile_size values leave visible seams along tile boundaries, while very large values inflate VRAM and can OOM on tight GPUs; tile_overlap and batch_size trade smoothness and throughput against memory.
CostThe selected checkpoint is pulled on first run and cached; the model choice is fixed at startup, so switching it requires a restart.
I/O contractAn empty input frame is returned unchanged without invoking the model; input is converted to RGB and output is always RGB regardless of source color space.
LimitationOutput height and width match the input — the component restores but does not super-resolve.

Versions

82162511latestdefault
Automated release
6/28/2026