Skip to content

QA Image Diff

Put your before and after side by side for the computer. It draws a map of what changed and gives you numbers, so you can check that a retouch only touched what you meant it to.

QA Image Diff — before / after
In ComfyUI Add Node → ComfyUI-JH-PixelPro → qa → QA: Image Diff (SSIM/PSNR)
Good for Checking where a retouch changed pixels, comparing two settings or pipelines, verifying an edit stayed inside a mask

When to use it

Our eyes miss subtle changes. This node measures them:

  • Diff image: shows where the two images differ. Bright means a bigger change.
  • SSIM (structural similarity): from about 0 to 1. 1 means identical structure. It reacts to changes in texture and edges more than to a small overall brightness shift.
  • PSNR (in dB): higher means closer. Identical images give the cap, 100 dB.
  • Report: JSON with ssim, psnr, mae (average absolute difference, 0 to 1), max_abs (largest single difference) and changed_pct (percentage of pixels that moved by more than 1/255).

Use it to prove that a frequency-separation or smoothing pass left the eyes and hair untouched, or to see exactly which areas a new setting changes.

Quick start

  1. Connect the original to image_a and the retouched version to image_b. Both must have the same width and height.
  2. Leave diff_mode on heatmap and amplify on 4.
  3. Connect diff to a Preview Image node. Connect report to a text preview node to read the numbers.
  4. Optionally connect a mask to measure only one area.

Settings

Setting Default What it does Tip
diff_mode heatmap heatmap is a colour map of the strongest channel difference. abs is the per-channel absolute difference. ssim_map shows where the structure changed (bright = changed). Start with heatmap.
amplify 4.0 Gain applied to the difference so small changes show up. Not used by ssim_map. 1 to 100. Raise it to reveal subtle edits.
window_size 11 Size of the SSIM window (odd number). Larger = judges coarser structure. 3 to 31. Reduced automatically on small images.
mask (optional) none Restricts all the numbers to this region, weighted by the mask. The diff picture itself is not masked.

Outputs: diff (IMAGE), ssim (FLOAT), psnr (FLOAT), report (STRING).

Batches

The ssim and psnr outputs refer to the first image of a batch only. The report lists every item.

Tips from retouching

  • A very high amplify makes tiny rounding differences light up. Compare against the numbers, not only the picture.
  • Compare against the mask you used for the edit. If changed_pct inside the mask is high but outside is near zero, the edit stayed in place.
  • If the mask is entirely black, the numbers return neutral values (SSIM 1, PSNR 100, others 0).
  • Use ssim_map to ignore uniform brightness shifts and spot only structural changes.
Under the hood

SSIM is computed with kornia.metrics.ssim over the chosen window and averaged over channels. PSNR is -10 log10(MSE), capped at 100 dB. heatmap applies the Kornia turbo colormap to the maximum channel difference times amplify. With a mask, every metric becomes a mask-weighted mean. Images must match in size (there is no automatic resize); a batch of 1 is broadcast. Source: core/qa_diff.py.

Example workflow

QA Image Diff example workflow. Copy workflows/sample_portrait.jpg into ComfyUI/input/ first.