Batch Images

This is the modern "Batch Images" - the node ComfyUI actually wants you to use when you need to stack several images into one batch. If you've loaded a recent default workflow or searched for "combine images", this is what you found. It replaces the old two-input ImageBatch node, which still exists but walks around with "(DEPRECATED)" permanently stapled to its name.

The pitch is simple: plug in as many images as you want, get one batched IMAGE out. The input ports auto-grow as you connect more wires - the schema defines a single autogrow input with slots for 1 to 50 images - so there's no fiddling with "add another input" menus. It just keeps making room.

Why batching matters

Remember that an image in ComfyUI is really a tensor shaped [batch, height, width, channels]. Batching is stacking images on that first axis so downstream nodes process them together. The payoff is two-fold:

• Speed. A sampler sees a batch of 8 and runs all 8 denoising passes while the model is hot in VRAM, instead of reloading weights 8 times.
• Consistency. Everything downstream - same seed, same settings - sees one logical run.

Batch Images is the node that makes "I generated 8 seeds, now upscale them all" a single graph instead of eight.

What it does, mechanically

Concatenation along the batch axis. Pure, boring torch.cat. It makes no attempt to reconcile anything - and that's the one real behavioral difference from its predecessor.

The old ImageBatch node would silently upscale the second image if the sizes mismatched and pad alpha channels to make things fit. This node does none of that. Every image you feed it must have the same height, width, and channel count, or the batch construction fails. That's a feature, honestly: a loud error beats a quietly-resampled image any day. If your inputs genuinely differ in size, resize or crop them first (there are core nodes for both), then batch.

The inputs that matter

• images - the autogrow input. Wire in between 1 and 50 IMAGE tensors. Order matters: the batch stacks in connection order, so batch position 0 is the first wire you plugged in. If your downstream node cares about order (say, you're matching frames to captions), pay attention to the order you connect.
• Output: one IMAGE with batch size equal to the number of inputs.

That output can go anywhere an image goes: into a KSampler, an upscaler, a VAE Encode, a Save Image, or another Batch Images if you're being perverse about it.

When you'd reach for it

• Gathering several individually-generated images (different seeds, different prompts) into one batch for a shared post-process pass.
• Stacking frames from Load Image nodes into one tensor before a video or animation step.
• Replacing a chain of old ImageBatch nodes in a legacy workflow - this is the drop-in that does the same job with one node and no resize surprises.

The search aliases tell the story: "combine images", "merge images", "stack images". If you're trying to do any of those three things with more than two images, this is the node.

Common issues

• Size mismatch errors. Every input must share height, width and channels. The error message is unhelpfully terse, so if it fails, check that you didn't accidentally wire a 512px image next to a 768px one.
• Order surprises. The batch order is wire order. If you assumed "first connected = first in batch" and got something shuffled, check which port you plugged into when.
• Don't expect resizing. Unlike the old node, mismatched images will error rather than resample. Route through a resize node first.

Ships with ComfyUI core - no install, no models, no keys. One node, up to 50 images, all the way to one batch.