🅛🅣🅧 LTXV Add Latent Guide

LTXV Add Latent Guide is the keyframe-anchor node for when your reference is already a latent, not an image. Its description says it plainly: it adds a keyframe or a video segment at a specific frame index. The difference from the image-based Add Guide nodes is what it eats - a LATENT you've already encoded - which makes it the natural pick when you're chaining generations, stitching segments, or extending a clip using latents you already have in hand.

It's part of ComfyUI-LTXVideo, Lightricks' official pack. Anchoring on a latent instead of an image skips a re-encode and keeps you in latent space, which is exactly what you want when the "keyframe" is the output of an earlier LTX pass rather than a fresh picture off disk.

How it works

You pass in the video latent you're building and a separate guiding_latent - the encoded segment you want anchored - plus the index where it should sit. The node injects the guiding latent into the conditioning and the target latent at that position, and the sampler generates a video that includes it. Because it operates on latents directly, it's the clean way to say "start (or end, or pivot) on this already-generated segment." A strength control governs how firmly the anchor holds.

The inputs that matter

• vae - present for the encode/decode plumbing the conditioning path needs.
• positive / negative - conditioning in, modified out.
• latent - the video latent you're conditioning.
• guiding_latent - the encoded keyframe or segment to anchor. This is the input that sets this node apart from the image-based guides.
• latent_idx - where it lands (default 0; negatives count from the end).
• strength - anchor hold (default 1, range 0–1).

Outputs are the updated positive, negative, and latent - send them to your sampler.

Installing it

ComfyUI Manager → Install Custom Nodes → search LTXVideo → install → restart, or cd ComfyUI/custom_nodes && git clone https://github.com/Lightricks/ComfyUI-LTXVideo and restart. It appears under ltxtricks. No weights of its own; the LTX checkpoint and, on LTX-2, the Gemma encoder are the downloads.

Where people get burned

• Feeding it an image. This node wants a guiding_latent. If you have a raw image, either VAE-encode it first or use the image-based LTXV Add Guide / Add Guide Advanced instead.
• Latent scale mismatch. The guiding latent has to match the target latent's spatial and temporal grid. A latent encoded at a different resolution or from a different VAE won't align - encode both through the same pipeline. Note LTX-2.3's rebuilt VAE means latents from an older LTX version aren't compatible.
• Index off the grid. LTX's temporal grid ((8n+1) frames on 2.3) constrains where an anchor can sensibly go; an arbitrary latent_idx can land between valid positions.
• Strength-induced freeze. As with the other guides, a full-strength anchor can lock the surrounding motion. If the segments around your anchor go static, ease strength down.