LTX2 Sampling Preview Override

Generating LTX-2 video without previews is a special kind of misery. A clip can take minutes, and if you're sampling blind you don't find out the motion went sideways or the face melted until the whole thing decodes. LTX2 Sampling Preview Override puts the live latent preview back, so you can watch the video take shape and kill a bad seed early.

The node is refreshingly honest about what it is - its own description calls it "a temporary measure until previews are in comfy core." LTX-2's new latent space didn't slot into ComfyUI's stock preview path (the same reason the old TinyVAE preview trick stopped working on LTX 2.3), so kijai wrote an override that decodes cheap previews during sampling. It's a stopgap, but a genuinely useful one while it's needed - LTX 2.x is finicky enough that seeing the latent evolve tells you within a few steps whether to cancel and reseed.

How it works

It swaps in its own preview function for LTX-2 sampling. Every so many steps, it decodes the in-progress latent into a rough frame and shows it in the node UI. You still get the full-quality decode at the end; this just gives you a running look. It patches your model and passes it straight through - the MODEL you get back goes into your sampler exactly where the original went.

The inputs and outputs that matter

• model (MODEL) in and MODEL out - drop it in the model line ahead of your sampler.
• preview_rate (default 8, range 1–60) - how often it refreshes the preview. Lower means more frequent previews (and a bit more overhead); raise it to preview less often and claw back a little speed.

Two optional inputs sharpen the previews: vae (VAE) lets it decode a proper-looking preview instead of a crude approximation, and latent_upscale_model (LATENT_UPSCALE_MODEL) is there for workflows that upscale in latent space. Neither is required to get a usable preview.

How to install it

ComfyUI Manager → search KJNodes for ComfyUI → install → restart. Or manually:

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

then restart. No model downloads tied to this node specifically - you're already loading LTX-2's weights elsewhere in the graph.

Common issues & troubleshooting

Previews cost a little time and VRAM. Decoding preview frames mid-sample isn't free. On a tight card, if you're right at the OOM edge, either raise preview_rate so it decodes less often or drop the optional vae input and accept rougher previews.

Previews look muddy. Expected without a vae wired in - the cheap approximation is meant to show you motion and composition, not final quality. Feed it your LTX-2 VAE for a cleaner look.

It doesn't do anything on a non-LTX model. This is LTX-2 specific by design; it won't add previews to a Wan or Flux graph.

One day you won't need it. The whole reason it exists is that core ComfyUI didn't have LTX-2 previews yet. If a ComfyUI update ships native previews for the model, you can pull this node out - that's the plan the author states outright.