LTXVMakeRefVideo

IC-LoRAs are where LTX gets its control: an in-context LoRA conditions generation on a reference video, not just a prompt. The catch is you usually have a handful of stills - a character sheet, a few product shots - and LTX wants frames. LTXVMakeRefVideo is the pack's answer: it takes a batch of reference images and expands them into a reference video with the right frame count for an IC-LoRA conditioning pass. No model, no inference - pure geometry, and it's fast.

How it works

Two inputs: images (the reference batch) and frame_count (default 17, minimum 17, stepping by 8). The output is a single IMAGE batch - the expanded reference sequence, ready to feed an LTX IC-LoRA conditioning node.

The mechanism is where the care went in. The node treats your last image as the background and every image before it as a subject: each subject is scaled (contain-fit, centered on a blank canvas) so it matches the background's dimensions, then the whole batch is laid out across frame_count frames. The subject images are distributed so each one occupies the frames that map to distinct latent positions - LTX's temporal compression means one latent covers 8 frames, so the node aligns subject placement to latent boundaries rather than raw frame counts. The last image repeats through to the end as the persistent background.

In plain terms: your subjects appear as reference frames spaced through the clip, and the background holds everything together - exactly the shape an IC-LoRA reference video needs to convey "this character, in this setting."

Why you'd use it

Any LTX 2.x IC-LoRA workflow (character reference, subject consistency, control-video tasks) where you're starting from stills instead of a real clip. It slots right before the IC-LoRA encoder: make the reference video here, feed it downstream, generate. It also pairs naturally with LTXVAddGuidesFromBatchIndexes in the same pack - build a reference sequence here, stamp guides into the latent there.

Install

It's in ComfyUI-Easy-Media:

cd ComfyUI/custom_nodes
git clone https://github.com/yolain/ComfyUI-Easy-Media.git

restart ComfyUI (ComfyUI Manager, "Easy-Media"). No models needed - it's tensor reshaping, not inference. You'll obviously want an LTX IC-LoRA and the LTX ComfyUI nodes on hand to make the output useful.

The gotchas

Frame count matters: 17 is the default and honestly a good starting point for a short reference clip - you can stretch to 25/33 if your model handles longer references. The background/subject convention is the thing to respect: if you accidentally put your background first instead of last, your subjects get treated as backgrounds and vice versa. And every input image needs the same channel count, or the node errors out - normalize your images before you build the batch. For a utility node that does no AI at all, it's remarkably easy to get wrong in the fun ways.