Image Blend (CCN)

Image Blend does one thing and does it cleanly: result = image_a * (1 - blend) + image_b * blend. Two images in, a weighted mix out, with a blend slider from 0 (all A) to 1 (all B). In the source's own words, it exists for the classic upscaling problem: you upscale a video or an image, and the model shifts the color slightly. Blend the upscaled result with the original and you keep the new detail while the hue stays honest.

That color-preservation use case is where it earns its keep. Upscalers are trained to add detail, not to be loyal to your palette, and a lot of "why does everything look warmer after upscale" complaints trace straight back to that. One Image Blend between the original and the upscaled pass at blend 0.3–0.5, and you've got an honest compromise: sharpness and texture from the upscaler, color from the source.

The inputs that matter

• image_a and image_b - the two image sets. They don't need to match, because the node handles the awkward cases for you.
• blend - 0 to 1, default 0.5, step 0.01. The whole node.

The part that makes it actually usable

Batch and resolution mismatches are silently handled instead of erroring out. If the two inputs have different batch sizes, it truncates the larger to match (with a console note). If resolutions differ, it resizes image_b to image_a's resolution before blending. That matters more than it sounds: image sets from different stages of a pipeline almost never arrive at the same shape, and a node that throws "shape mismatch" at you for that is the status quo. This one just works, which is the whole reason to reach for it over doing the math by hand.

The output is a single image - same batch as the inputs - ready to feed a VAE decode, a save node, or another compositing step. Nothing else is touched; it's a pure blend.

Install

It's part of ComfyCollectorNodes, so the standard one-liner applies:

cd ComfyUI/custom_nodes
git clone https://github.com/valkymaera/ComfyCollectorNodes

restart ComfyUI, and it's under ComfyCollectorNodes/Image. ComfyUI Manager also has it via ComfyCollectorNodes. No dependencies, no model files.

The honest take

Image Blend is the kind of node you don't think about until a workflow needs it, and then it's exactly right. It won't impress anyone, but it removes a real annoyance - most image nodes in this space expect perfectly matched inputs and throw errors when they aren't, while this one shrugs and normalizes. The one thing to remember is that blend is a plain weighted average: at high blend values the upscaled version dominates and color drift creeps back, so for color preservation you want to stay on the low side and let the detail ride on top. If you're doing a lot of this, it pairs nicely with the pack's Latent Clamp and Normalize nodes on the decode side, but that's a separate rabbit hole.