Composite Image Splitter v5.1.0

Sometimes a workflow hands you one image that's really two images glued together - a before/after, a source-on-the-left/target-on-the-right comparison, or the "stacked images for cat vton with flux" composites this pack loves. tri3d-composite-image-splitter cuts every image in a batch exactly in half, vertically, and returns the two halves as two outputs. That's the whole node, and it's surprisingly handy once you realize how often your input is a side-by-side.

What it does

One input - images, an IMAGE batch. For each image it finds the width w, takes the left half ([:, :w//2, :]) as one output and the right half ([:, w//2:, :]) as the other, and stacks the halves back into two full batches.

• image1 - the left halves
• image2 - the right halves

No options, no crop parameters, no resizing - the halves come out at the original height and half the width, preserving the resolution of each side. If your composite is uneven (text on one side, subject on the other), the split is still straight down the middle; there's no smart content detection.

Why you'd use it

Three common cases: you generated or downloaded a side-by-side comparison and want each side separately; your VTON pipeline feeds a "model | garment" stacked image and you need the halves for different processing paths; or you're splitting a training/comparison composite so each half can go through its own branch. The batch handling is the quiet win - a folder of composites in, two clean stacks out, aligned by index.

Installing

ComfyUI Manager (search "tri3d"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/TRI3D-LC/tri3d-comfyui-nodes

Restart, install requirements (pip install -r custom_nodes/tri3d-comfyui-nodes/requirements.txt). No models, no downloads - a slice plus a stack.

Troubleshooting

• Halves look squished - they aren't; a half of a side-by-side has the same height but half the width, so it's correctly proportioned for the content, not for the original frame. If the original was stacked vertically instead of side by side, this node splits the wrong axis - that's a different job.
• Content is cut in the middle - expected, it splits blind at the midpoint. Check your composite is actually 50/50.
• Batch size changed - both outputs keep the same batch length as the input; if you expected fewer, you may have fed it a single stacked image where you meant to feed a list.

A one-trick node, but the trick - blind, batched, middle-of-the-image splitting - is exactly what the pack's composite-based workflows need.