Academia SD LTXV Multi-Frames 🖼️

LTX-Video's image-to-video is genuinely fast - that's its whole personality - but wiring multiple reference frames into it is normally a rat's nest of encode nodes, concatenation, and latent mask fiddling. This node collapses all of that into one box: drag a few images in, set a strength per frame, and it injects them directly into the latent space at the right indices. No extra Load Image nodes cluttering the canvas, no manual latent math.

How it works

The mechanism is honest about its lineage: the README credits Kijai's ComfyUI-LTXVideo wrapper - specifically the LTXVImgToVideoInplaceKJ node - as the base, adapted into a cable-free UI. You feed it a vae and a latent (the base latent from an Empty Latent node sized for your clip), plus kf_data, a JSON list the node's own UI maintains with your images, their frame indices, and per-frame strengths. It encodes each reference image into the latent space, places it at its frame index (first frame at index 0, new frames default to -1 = last), and writes the noise mask so LTX treats those frames as fixed conditioning instead of noise to regenerate.

The kf_data string is the hidden state the UI edits for you - you'll never hand-write it. All your interaction is drag-and-drop and strength sliders. The output is a single latent you feed straight into your LTX sampler.

Why the indexing detail matters

LTX loops well when the first and last frames match, and the node's "first frame at 0, new frames at -1" convention is exactly what keeps a loop mathematically sound - you set your first frame, set your last frame, and the model animates between them without the seam. If you've ever watched an LTX clip snap back to a mismatched first frame at the end of the loop, you know what this saves you.

Install

Pack install: ComfyUI Manager → "AcademiaSD", or git clone https://github.com/AcademiaSD/comfyui_AcademiaSD, restart. Important: this node doesn't install Kijai's ComfyUI-LTXVideo for you, and while it reimplements the injection logic in-repo, you still need an LTX-Video checkpoint plus ComfyUI's native LTX support in your environment for the sampler side of the workflow. If you see "no module" errors mentioning ltx, the model files are the missing piece, not this node.

The honest take

If you do LTX image-to-video at all, this is a real quality-of-life upgrade - the per-frame strength control alone beats the stock single-image workflow. The one thing to keep in mind is that LTX is a fast but finicky model: multi-keyframe setups multiply the ways it can drift, so expect to iterate. The node makes the mechanics painless; the model's temperament is still your problem.