🖼️AGSoft Load Image & Mask Resize

The classic way to quietly ruin an inpainting workflow is to upscale the image with one node, upscale the mask with another, and hope they still line up. Two independent interpolation passes, and by the time you've thrown in a crop or pad you're a few pixels off - which shows up as seams, or as the sampler regenerating a region you meant to keep. AGSoft Load Image & Mask Resize is the "do it in one place" version: it loads an image and its mask, then runs rotate → resize → flip over both as a single synchronized unit. They can't drift apart because they share one code path.

You'd reach for it at the front of an inpaint or outpaint pipeline, when you need to get a source image onto a model's native resolution and into a latent-friendly shape before anything else runs. It's also the right tool when the resize has conditions on it - "only shrink if it's bigger than the target," say - because those live in the same node instead of a hand-rolled If block.

Sources, in priority order

• input_image (optional) - an image tensor from another node, highest priority.
• custom_path - absolute path or linked string, next in line, validated at connect time.
• image - the combo/upload widget, used when neither of the above is set.

That first option is what separates this from the plain load node: you can chain it after any node that emits an IMAGE and never touch disk.

The inputs that actually matter

The node has a lot of knobs; these are the ones a beginner sets:

• resize_mode - seven ways to express the target: target_percentage, target_megapixels, target_width, target_height, target_both, target_longest, target_shortest. Megapixels mode is genuinely handy when you care about total pixel budget (upscaling) rather than exact dimensions.
• fit_mode - this owns how it resizes: stretch (distort to the exact target), crop (scale proportionally, cut the overflow), pad (scale to fit, fill bars), proportional (fit inside the target and output the resulting size - no bars, no crop).
• resize_if_larger / resize_if_smaller - pure triggers that decide whether to resize at all, independent of fit_mode. Set to none means always resize; otherwise the original size is kept when the condition doesn't fire. Pair these with fit_mode and you get conditional resizing without extra nodes.
• crop_position - nine anchors for where crop/pad cuts or pads.
• divisible_by - rounds the output dimensions to a multiple. Set 16 for SDXL, 8 for Flux; 0 is off. This is the one that stops the "latent size mismatch" error before it happens.

Worth knowing: interpolation defaults to lanczos, and for RGBA images the node auto-applies premultiplied alpha so resizing transparency doesn't give you that ugly dark halo. Rotation via angle_degrees (or the rotation_preset quick angles, which override the degrees field) runs before resize; flip_mode after. There's invert_mask if your mask convention is backwards, and device toggles CPU/CUDA - leave it on CPU unless you're doing huge batches, resizing is not the bottleneck.

Outputs are the same six as the plain loader: image, mask, width, height, filename, filepath - so it slots into any workflow that expected a load node.

Install

Same pack as everything AGSoft - the ComfyUI Manager route or the manual clone:

# ComfyUI Manager: search "comfyui-AGSoft" and install
# or:
cd ComfyUI/custom_nodes
git clone https://github.com/Art-xmaster/comfyui-AGSoft.git
# restart ComfyUI

Dependencies are light (numpy, opencv-python, translators) and this node needs no model files. Find it under AGSoft/Image in the node menu.

Gotchas

• A custom_path typo fails at connect time, not at run - that's the node doing its job, but it means a string from another node that's empty will fall back to the combo rather than error. Check your wiring if the "wrong" image loads.
• Mask too clean to be true? If nothing is connected and the source has no alpha, the mask output is all zeros. Fine for a plain resize, useless for inpainting - connect a real mask or feed it a transparent PNG.
• The README is Russian-first and detailed node docs live on the author's Telegram channel; the in-node tooltips are bilingual, so start there.