Image Patch Editor

The Image Patch Editor does one thing, and it's the thing every whole-frame editing model is secretly bad at: leaving the rest of the image alone. Feed it an original and an edited version of the same image, and it figures out which pixels actually changed, builds a mask by itself, and composites the edited content back onto the untouched original. No mask painting, no bbox dragging. That's the entire pitch, and it's a good one.

If you've fought Qwen-Image-Edit or Flux 2 Klein long enough, you know the problem: they take the whole image in and emit a whole new image out, so the stuff you didn't ask to change comes back close but not identical. Do it five times in a row and your character's face has subtly melted. The standard 2026 fix is crop-edit-stitch - render only the changed region, composite it back, leave every other pixel bit-identical. This node is that pattern minus the manual mask step: it detects the changed region for you. Reach for it anywhere you'd otherwise hand-draw a mask and composite with ImageCompositeMasked.

How it works

Three passes, all CPU, all classic computer vision. First it can auto-align the edited image to the original using ORB feature matching with a RANSAC fit - the README's claim is real, and worth knowing the limits of: the transform only handles uniform scale and translation, never rotation or shear. Align a cropped-and-upscaled region and it's perfect; rotate the edit and it silently falls back to a plain resize.

Then it computes a difference heatmap with one of three algorithms: difference (plain pixel-wise absolute difference), l2_norm (Euclidean color distance - more sensitive to hue shifts), or ssim (perceptual structural similarity, slowest but most human-like). That heatmap gets thresholded into a binary raw_mask, and then a pipeline of cleanup: min_area_size drops small specks, mask_expansion dilates or erodes it, blur_radius feathers the edges. Finally the edited image is composited onto the original with Image.composite using that mask.

The inputs that matter

Only a few actually need touching from the defaults:

• tolerance - the one to tune, and it's counterintuitive: higher detects more changes, lower only catches dramatic ones. Start at 0.3 and crawl up until the mask covers what changed and nothing else.
• min_area_size - the noise gate. Raise it when the heatmap is picking up JPEG grain and compression artifacts as "changes."
• detection_alg - ssim for quality, difference or l2_norm when you want speed.
• region_size / stride - the sliding window. Keep stride at 1 for full overlap (slowest, most accurate); bump it toward region_size for a fast approximation.
• edit_max_alpha - cap how strongly the edited pixels come through; useful for fusing two model outputs without hard cutovers.

The outputs you'll actually wire up: merged_image goes straight into a Save Image node, final_mask is a proper MASK you can feed anywhere else, and debug_overlay - original with the detected changes highlighted in red - is your sanity check for that first run. heatmap and raw_mask are for tuning and curiosity.

Installation

No models, no API keys, no VRAM - it's pure image math. The heavy dependency is scikit-image, which the code genuinely uses (ORB, RANSAC, SSIM).

cd ComfyUI/custom_nodes
git clone https://github.com/Premik/ComfyUI-ImgPatchEditor
cd ComfyUI-ImgPatchEditor
pip install -r requirements.txt

Restart ComfyUI and look for Image Patch Editor under the ImgPatch category. ComfyUI Manager also finds it by searching "Image Patch Editor" or "ImgPatchEditor". (Source note: the README also mentions a pillow-avif-plugin dependency, but it's commented out in requirements.txt - you only need scikit-image and numpy, and ComfyUI already ships Pillow.)

Gotchas

The node requires original_image to be a single image - it raises a clear error if you feed it a batch. edited_image, though, accepts a whole batch, and it processes them chained: each edited frame patches the output of the previous one. That's the intended iterative-refinement workflow, and also a footgun if you expected independent outputs. And one more trap: if the sizes mismatch and allow_rescale is off, you get a hard ValueError - turn allow_rescale on to let the ORB alignment sort it out.

It's a young, single-node pack with zero community footprint yet, so treat the defaults as a starting point and trust debug_overlay, not your eyes, when tuning. For the "edit a little, keep the rest pristine" crowd this is quietly the most useful three lines of a workflow you'll add this month.