Compairy-PhotoDeblur: Native nodes for photorealistic elimination of the back and hair of the camera without loss of resolution. All this is based on special models HI-Diff and MISCFiter. Unlike traditional graphic materials, which often finish fictional details or smooth plastic, these points focus on the correct physical recovery of sharpness. The tool solves the typical problems of shooting handheld in low light: light drops around night marks, blurred window borders, loss of sharpness on eyelashes, hair, cloth and repeating architectural textures. The architecture supports two models: diffusion HI-Diff (step by step sampling with an ordered expression of microrelief) and direct filter MISCFiter, which works in one pass of the network without diffusion steps. To process multi-megapixel images introduced tile processing (driven processing) with custom overlap (default Harry 768, overlap 96) and smooth seams through linear, cosine or Gauss mixing, which prevents overflow of VRAM video memory. NodePipeline also built in TTA, 8-round averaging and mirror projections to suppress double contours, and the Match Input Tone module, which preserves the frame's original color balance. The official barriers trained on RealBlur J, RealBlur R and GoPro Reference datasets are automatically downloaded and validated using the built-in load node. The terms: @cgevent.

Compairy-PhotoDeblur: Native nodes for photorealistic elimination of the back and hair of the camera without loss of resolution. All this is based on special models HI-Diff and MISCFiter. Unlike traditional graphic materials, which often finish fictional details or smooth plastic, these points focus on the correct physical recovery of sharpness. The tool solves the typical problems of shooting handheld in low light: light drops around night marks, blurred window borders, loss of sharpness on eyelashes, hair, cloth and repeating architectural textures. The architecture supports two models: diffusion HI-Diff (step by step sampling with an ordered expression of microrelief) and direct filter MISCFiter, which works in one pass of the network without diffusion steps. To process multi-megapixel images introduced tile processing (driven processing) with custom overlap (default Harry 768, overlap 96) and smooth seams through linear, cosine or Gauss mixing, which prevents overflow of VRAM video memory. NodePipeline also built in TTA, 8-round averaging and mirror projections to suppress double contours, and the Match Input Tone module, which preserves the frame's original color balance. The official barriers trained on RealBlur J, RealBlur R and GoPro Reference datasets are automatically downloaded and validated using the built-in load node. The terms: @cgevent.