LTX Reference Images

This is the boring node in the pack, and it's boring on purpose. LTX Reference Images does exactly one thing: it collects up to six reference images into one ordered list so the Multi Reference Slots node can consume them as a batch. No model, no VAE, no math - six inputs, one list out. You'll still want it, because it solves a genuinely annoying problem.

The problem is that a plain IMAGE batch in ComfyUI can't carry images of different sizes - torch stacking needs a single shape. And subject references are rarely the same size. A portrait crop of one character next to a landscape product shot of another simply won't stack into one tensor. This node emits a list instead, so each reference keeps its own resolution, and your slot order lives in one visible place instead of being scattered across six inputs on the conditioning node.

The inputs

image_1 through image_6, all optional. Leave one unconnected and it's skipped - and here's the subtle part from the tooltips: connected slots renumber in order. Wire image_1 and image_3 and you've got references 1 and 2, which the prompt calls <Image 1> and <Image 2>. The numbering that matters is the connected order, not the input slot number, so keep that straight when you're building the tags.

The outputs

• references (LTX_REFS) - the list, ready to plug into the references input on LTX Multi Reference Slots. That node takes the list or the individual reference_N slots; if both are connected, the list wins, so you don't need to double-wire.
• count (INT) - how many references actually landed. Handy as a sanity check, and if it's 0, you know nothing got connected before you waste a generation wondering why.

Install and troubleshooting

It's part of ComfyUI-BFSNodes, so install the pack once: ComfyUI Manager → search "ComfyUI-BFSNodes", or git clone https://github.com/alisson-anjos/ComfyUI-BFSNodes.git into ComfyUI/custom_nodes and restart. This node itself needs nothing beyond torch - the pack's heavier requirements (insightface, onnxruntime) are for its face-loader nodes, not this one.

There's not much to break here. The one real failure mode is wiring an image into reference_N on the slots node and into image_N here and expecting them to merge - they don't, and the list silently overrides the individual slots. Pick one path. And remember the collector is only half the workflow: your prompt still has to name the slots via LTX Reference Tags, because the model has no idea what <Image 1> is unless the training convention tells it.