Multi-Load Cowboy

Any workflow that needs a handful of reference images starts with a sprawl: six Load Image nodes, six resizes, a dozen wires. This node collapses it into one node whose face is a grid of slots sitting next to the output column - so all eleven outputs cost no extra height. Click a slot to browse the input folder or upload, drop files straight onto slots, drag one slot onto another to swap, and clear one with its corner X. It's the most genuinely well-thought-out loader UI in the pack.

How the slots behave

The six required image_1..image_6 inputs are enum selectors over your input folder, defaulting to (empty). Empty slots are skipped, not errors - a half-filled grid is a valid graph, and an all-empty grid returns a black frame plus count of 0, so a fully empty node won't crash a template. Drop several files at once and they fill the empty slots in order; drag a slot onto another to swap them.

The shared resize

Every loaded image is resized to one common size - that's the point of "one shared resize." width and height (default 1024×1024; 0 on either means derive it from the other), resize_mode, upscale_method, crop_position, pad_color, and divisible_by (rounds the output down to a multiple, 8 or 16 for most models).

The resize modes follow KJ's Resize Image V2 vocabulary, so if you've used that pack it'll feel familiar:

• stretch - fills the box, changes aspect ratio
• resize - fits inside the box, keeps aspect (can be smaller than the box)
• pad - fits inside, fills empty space with pad_color
• pad_edge - fits inside, stretches edge pixels into the gaps
• crop - fills the box exactly, cuts off what doesn't fit
• total_pixels - keeps aspect ratio, area matches width × height

upscale_method defaults to lanczos - sharpest, CPU-only. For a big batch on GPU, switch to bicubic or bilinear; nearest-exact for pixel art. crop_position picks which part survives a crop and where the image sits in a pad.

The outputs

The batch output (images) holds all filled slots padded up to a common size when a fit mode leaves them uneven. The masks output carries each slot's alpha with padded regions reading as 1.0 - so when you pad a slot, the mask tells downstream nodes where the real pixels end. Plus count, width, height, and six individual image_1..image_6 outputs for wiring separate branches - because sometimes you want slot 3 on its own and slots 1-2 batched.

Installing it

Part of TrentNodes:

cd ComfyUI/custom_nodes
git clone https://github.com/TrentHunter82/TrentNodes.git
cd TrentNodes
pip install -r requirements.txt

ComfyUI Manager ("Trent Nodes") also works, modulo the standing caveat: the author renamed the repo on day one, leaving two registry entries that occasionally make Manager flag the pack as "unsafe" - clone manually if it refuses. No model downloads; OpenCV does the heavy lifting.

Where people get burned

• Expecting empty slots to be an error - they're not, by design. If count is 0 and your output is a black frame, that's the node telling you no slots are filled, not a bug.
• Lanczos slowness - it's CPU-only and can crawl on large batches. If a big grid feels sluggish, that's why; switch to bicubic on the GPU.