Lazy Mix Images | akatz-loops

LazyMixImages | akatz-loops blends two images by a mask, but it's built for the lazy-execution world, and that changes how you should think about it. It takes image1, image2, and a mask, and does the classic image1 * (1 - mask) + image2 * mask composite. What makes it interesting: it inspects the mask first, and if the mask is all zeros it never evaluates the image2 branch at all - and vice versa. The image that's fully masked out isn't just ignored, it's not even computed.

Inputs and outputs

• image1 (IMAGE) and image2 (IMAGE) - both lazy inputs.
• mask (MASK) - the blend weights, 0 = all image1, 1 = all image2, in-between = the weighted mix.
• output - a single IMAGE.

Two nice touches in the implementation: it handles 2D masks (adds the batch/channel dims it needs), and it broadcasts a single-channel mask up to match the image's channels. So you can feed a plain black-and-white mask from a mask-utility pack and it just works.

When you'd use it

The lazy angle shines inside a loop. Say each iteration either needs a clean pass-through or a mix with a freshly generated frame, decided by a mask that's sometimes fully black. A normal blend node would force both upstream image branches to render every iteration. This one only renders what the mask actually needs - which, over a 30-iteration loop, is a real time save.

It's also simply a clean composite node for "switch by mask" UI work - an inpainting result over the original, a region overlay - when you want the readability of one node instead of a Composite plus a Blend tangle.

Installing it

Part of the Akatz-Loop-Nodes pack (repo ComfyUI-Execution-Inversion):

cd ComfyUI/custom_nodes
git clone https://github.com/akatz-ai/ComfyUI-Execution-Inversion
# restart ComfyUI

Or ComfyUI Manager → "Akatz-Loop-Nodes". Only dependency is opencv-python; no model files.

Gotchas

The author's own comment flags it: it does not handle different batch sizes between the two images. If image1 is a batch of 4 and image2 is a single frame, you'll get a shape mismatch, not a broadcast. Keep both images (or both masks) at the same batch size. Also, the mask check looks at the whole mask's min/max - a mask with a single stray 1.0 pixel means image2 will be evaluated, so the lazy shortcut only pays off for genuinely solid masks. Not a bug, just set expectations.