Concat Multiband

Concat Multiband is the boring, dependable node in the pack: it takes two MULTIBAND_IMAGE objects and joins their channels into one. No math, no resizing, no cleverness - it's torch.cat along the channel axis with a nicer wire. If Compose Multiband is how you build a stack from loose masks, this is how you build it from two stacks you already made.

The honest use case is modularity. Build one multiband that holds your segmentation masks and another that holds, say, feature maps or depth channels. Rather than rebuilding a single 12-channel stack whenever either side changes, you keep them as two sub-stacks and join them with one node at the end. Re-run only what changed; the concat is cheap.

How it works

multiband_1 and multiband_2 each arrive as (B, C, H, W) tensors with their channel-name lists. The node concatenates along the channel dimension and concatenates the name lists to match: 5 channels + 3 channels = 8 channels, names preserved in order. Metadata dicts from both sides get merged too, so anything the load/save round-trip carries survives.

Same constraints as the rest of the pack: batch sizes must match, and spatial dimensions must match. It raises a clear "Spatial size mismatch" error rather than guessing, so you'll know. There's no output limit - join a 4-channel image stack with a 30-channel spectral stack and you get 34 channels out.

The inputs

Only two, both required:

• multiband_1 / multiband_2 - any two MULTIBAND_IMAGEs. Order matters for channel order downstream: whatever you want first on the wire goes first.

Output is a single multiband - same type in, same type out, so you can chain Concat into Concat or feed the result straight into Save Multiband Image, Select Multiband Channels, or Multiband to Image.

Installing the pack

Part of ComfyUI-Multiband. ComfyUI Manager: search "ComfyUI-Multiband" → Install → restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/PozzettiAndrea/ComfyUI-Multiband.git
pip install -r ComfyUI-Multiband/requirements.txt

Dependencies are just numpy, torch, and tifffile; no model files to fetch, since the whole pack is tensor manipulation rather than inference.

One thing worth knowing

Concat is the pure channel-join. If your two stacks have different spatial sizes and you want them in one multiband anyway, run Resize Multiband on the smaller one first, or use Compose Multiband - no, actually Compose enforces the same match. There's no forgiving path here: get the sizes aligned, then concat. It's the price of a format that assumes every channel is the same resolution, which for this pack's use (segmentation masks, feature maps) is exactly what you want.