🖼️AGSoft Add Images

Five branches, each ending in an IMAGE, and the node you want to feed takes exactly one. Core's ImageBatch will glue two together - only two, and only if they're the same size. Stack three of them and you've got a little ladder of nodes whose ordering you'll forget by Thursday.

AGSoft Add Images is the blunt fix: up to 50 optional IMAGE inputs on one node, and one images output you can wire anywhere.

Why it exists, beyond "more inputs"

Because it makes a decision core nodes don't. resize_mode (default keep) tells the node how to reconcile mismatched sizes:

• keep doesn't touch your pixels. If every image is the same size you get a normal batch tensor - completely compatible with Preview, Save Image, and anything else expecting a tensor. If the sizes differ, you get a list of tensors at their original sizes.
• stretch forces everything to the first image's dimensions, distortion and all.
• contain fits proportionally inside the first image's size, padding the leftovers black.
• cover scales proportionally and center-crops to the first image's size.

The last three always give you one batch tensor, with upscale_method (default lanczos) picking the interpolation. The honest take: keep is the mode you'll live in, and its list output is the reason to install this at all. It's also the thing that will confuse you first, because a list is not a batch - see below.

How it works, mechanically

The node declares 50 optional image_1 … image_50 sockets. Unconnected optional inputs simply arrive as None and get skipped - the same convention every switch and fallback node in ComfyUI runs on. So the node walks the slots in numerical order, and that order is your output order, regardless of where the sockets sit on your canvas. Rewire a branch and you may silently reorder the batch.

Batches get exploded: if you hand it a tensor with 4 frames, those 4 frames enter the aggregation as 4 separate images. Channel counts are unified to the max (an RGBA image next to RGB ones pads the RGB side with alpha = 1, i.e. opaque). One thing worth knowing: the input side of the slot count is JavaScript, shipped in the pack's web/ folder. It keeps exactly (connected + 1) slots on screen and trims the other 49 so the node doesn't look like a fire escape.

The bits you set

resize_mode is the only widget most people ever touch; upscale_method matters only when you're not in keep. Then wire images into the slots. The output is images - hand it to Save Image, a preview, an images-to-video node, or the pack's own 🖼️AGSoft Image Stitch Plus, whose input_images accepts single tensors, batches, and lists, which is the intended pairing.

Installing it

ComfyUI Manager, search comfyui-AGSoft, or:

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

Restart ComfyUI. The pack's requirements.txt pulls numpy>=1.26.4, opencv-python>=4.7.0.72 and translators (that last one is for the pack's translate nodes, not this one - it installs regardless). No model downloads. This is a big pack from a Russian-speaking author whose documentation lives mostly on his Telegram channel, and it has essentially no Reddit footprint at all - you're reading the code, not the crowd, when something goes wrong. If the node doesn't appear, it's almost always the environment rather than the node: conflicting pins across a stack of installed packs are the standard ComfyUI injury, because there's no isolation between them.

Where people get burned

The list trap. keep + mismatched sizes → a list, and plenty of nodes only understand tensors. If a downstream node errors or does something weird, that's your first suspect. Fix it by switching to contain/cover, or by resizing upstream.

Which image is "first". stretch, contain and cover all target the first connected slot, i.e. the lowest-numbered one - not the biggest, not the one you dragged in last.

Sketchy slots. If the browser-side extension didn't load, you'll see all 50 sockets instead of a tidy few. The node still runs. A hard refresh fixes the cosmetics.

Alpha edge cases. Add Images always pads channels up, never down. Mixing RGBA and RGB gives you RGBA, with an opaque alpha on the RGB ones. That's fine for saving, occasionally not fine for a model that wants exactly 3 channels.