Poisson Blend (Seamless Clone)

Anyone who's pasted a subject onto a new background knows the tell: a hard edge, a colour cast, a bright rim where the old photo's lighting bleeds through. ComfyUI has a dozen "paste and pray" compositing tricks, but the honest fix is Poisson image editing - the same math Photoshop's "seamless clone" uses. Poisson Blend wraps OpenCV's seamlessClone (Pérez et al., 2003) into a proper ComfyUI node, so you can clone one image into another through a mask and get a blend that actually matches the destination's tone.

It's the newest node in the Eric_Composer_Studio pack - you won't find it in the pack README yet, and it sits under its own Eric_Composer_Studio/compose category rather than the main one. Worth knowing, because the canvas composer nodes in this pack are for layout, and this is the one that makes the pasted pixels believe they belong.

How it works

Instead of just overlaying pixels, Poisson editing solves a small gradient field: the pasted region keeps its own gradients (shape, texture, internal shading) while its overall colour and brightness drift toward the destination's local tone. That's why a bright subject can land on a dim background without looking composited. The node gives you three modes, each mapping to an OpenCV flag:

• normal (NORMAL_CLONE) - full colour/luminance match. Best for opaque objects.
• mixed (MIXED_CLONE) - at each pixel the stronger gradient wins. Best for transparent, wispy or textured things: foliage, smoke, hair.
• monochrome (MONOCHROME_TRANSFER) - the source contributes only luminance; hue and saturation come from the destination.

Inputs and outputs

Three required: background (destination canvas; the output matches its size), source (the image with the object to clone), and source_mask (which pixels are in - anything above 0.5 counts; soft masks get binarised because OpenCV needs a hard mask).

The optional list is where the craftsmanship lives:

• mode - one of the three above.
• auto_fit - on by default; resizes source and mask to background size first. Off, a size mismatch is an error.
• center_x / center_y - where the mask centroid lands on the destination; -1 uses the source mask's own centroid (i.e. no movement).
• offset_x / offset_y - nudges applied after centering.
• mask_shrink_px - erodes the mask before cloning. The default of 1 is there for a reason (see below).
• post_feather_px - a feathered re-mix of the result over the background, to soften jagged mask edges without breaking the gradient match.
• strength - blends the Poisson result with the original background; 1.0 is full Poisson.

Outputs are image and used_mask (the mask actually applied, post-erosion - handy for debugging).

Installing it

Same as the pack: ComfyUI Manager → search Eric Composer Studio → install → restart, or git clone https://github.com/EricRollei/Eric_Composer_Studio.git into custom_nodes and pip install -r requirements.txt. The dependency that matters here is opencv-python, which is in the requirements.

Where people get burned

• Black borders / outright failure. OpenCV's seamlessClone throws a fit if the mask touches the image border. The node auto-pads when it detects that, but you should still keep mask_shrink_px >= 1 - it erodes away fringe artefacts at sharp matte edges too.
• "The paste didn't move." If you set center_x/y, remember they position the mask centroid and offset_x/y stack on top - and the node clamps so the mask's bounding box stays inside the destination. If your subject is already exactly where you want it, leave centering at -1.
• Source size mismatch. If auto_fit is off and dimensions differ, it's a hard error by design. Crop or pad the source to background size first, or just leave auto_fit on.

One honest caveat: Poisson blending is great at tone-matching, but it can't fix a subject whose geometry says "I was lit from the left" landing in a scene lit from the right. That's a lighting problem, not a seams problem - no solver in OpenCV will save you there.