AccioPix unblur lab, 10 September 2026 — https://acciopix.com/blog/how-to-unblur-a-photo

reference.jpg is the sharp photo (Pexels, Mat Reding, https://www.pexels.com/photo/brick-buildings-under-blue-sky-11363655/)
resized to 1600 px on the long edge. Each <case>-input.jpg is that photo blurred by a known amount:
  soft-1.5 / soft-3 / soft-6   Gaussian blur, sigma in px (missed focus, mild to severe)
  shake-8 / shake-16 / shake-32  linear motion blur, length in px at 15 degrees (camera shake)
<case>-nafnet.jpg   NAFNet (MIT), blind AI deblur, 512 px tiles (scripts/unblur-lab.py)
<case>-wiener.jpg   Wiener deconvolution with the TRUE kernel (an oracle: the best a classic deconvolver can do)
<case>-tool.jpg     the AccioPix unblur tool (https://acciopix.com/tools/unblur-image) run in a browser on the input,
                    Shake/motion mode for shake cases, Soft focus mode for soft cases
results-lab.csv     PSNR / SSIM against reference.jpg and variance-of-Laplacian sharpness for every row
The script that made all of this: scripts/unblur-lab.py in the AccioPix repository.