Deblur Image (FFTformer)

How it fits

Typical backends

Caveats

• Hard constraintOnly the single dynamic-scene motion-blur checkpoint is available; any other model selection aborts startup. It is not a denoiser, deraining, or defocus model.
• CostWhole-image inference is a single forward pass that can exhaust GPU memory on large frames; enabling patch-wise tiling switches to overlapped patch inference that bounds VRAM, with adjacent tiles blended to avoid seams. Even with tiling off, frames over roughly one megapixel auto-tile at a safe edge.
• FallbackA GPU out-of-memory error during inference is caught and the original RGB frame is returned unrestored rather than failing the stream; the restorer also falls back to CPU when no GPU is present, which is impractical for live video.
• State lifetimeThe dynamic-scene motion-blur checkpoint is supplied by the bound model file and loaded once at startup; there is no per-frame model reload.
• I/O contractAn empty input Image 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 sharpens but does not super-resolve.