🖼️AGSoft Image & Mask Resize

The difference between a useful mask node and an annoying one is whether the mask stays in sync with the image. AGSoft Image & Mask Resize is the lean version of the pack's resize family: feed it an image and a mask, tell it a size, and both come out resized together, with the mask keeping hard edges instead of turning into gray mush. It's the node you reach for when the fancy "Plus" version is overkill but you still don't want to resize image and mask separately.

How it works

One size spec, applied to both. resize_target picks the anchor axis - width or height. Then size is the target value on that axis (set to 0 to ignore it), and the other dimension scales to keep proportions. Leave size at 0 and scale_by takes over: 1.0 is unchanged, 0.5 halves, 2.0 doubles. Whichever way you go, the aspect ratio is preserved - no stretching.

Two details in the implementation matter in practice:

• divisible_by (default 8) - the final dimensions get rounded up to a multiple. Feed this into an SD or SDXL sampler and you sidestep the "not divisible" errors before they happen.
• Mask resampling is nearest-neighbor - while the image uses the upscale_method you picked (lanczos, bicubic, etc.), the mask is resized with nearest, so a crisp mask stays crisp. A blurred mask in a workflow is almost always someone resizing it with bilinear.

invert_mask flips black/white if your mask happens to be stored inverted, and device lets you pin the work to CPU or CUDA (it auto-falls back to CPU when CUDA isn't available).

Inputs and outputs

Required inputs: image, mask, resize_target, size, scale_by, upscale_method, invert_mask, divisible_by, device. Outputs: the resized image, the resized mask, and width/height ints for downstream nodes. Batches are handled - a single mask gets repeated to match a batch of images.

Where you'd use it

Classic case: you have a source image and a hand-painted mask at a different resolution (say the mask was drawn at preview size), and you need both at a common size before sampling. Also the standard pre-step for inpainting: resize the pair to a model-friendly resolution so the mask lines up with what the sampler actually sees. And in image-to-image work where you want a smaller latent but a full-res mask for the final composite - this keeps them matched.

Installing it

Pack standard: ComfyUI Manager → search comfyui-AGSoft → install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Art-xmaster/comfyui-AGSoft.git

No models, no dependencies beyond ComfyUI's built-ins.

Common issues

The two things that bite: forgetting that size beats scale_by when both are set (it's size-first logic), and expecting bilinear-smooth masks. If your mask comes back blocky, that's intentional - nearest is the right choice for mask edges, and you can soften it downstream with a blur if a specific model needs it. Also remember rounding is up to the divisible_by multiple, so a 500px target at divisible_by 8 becomes 504, not 496 - worth knowing if you're matching an exact pixel budget from another node.