Add LTX Latent Guide

LTXV Add Guide is how you tell LTX "this exact image should appear at this exact moment." Give it a picture and a frame index and it becomes a keyframe the generated video is built around - frame 0 for image-to-video, the last frame to control where a clip ends up, or a midpoint to force a beat you want. It's the backbone of keyframe conditioning in the LTX line, and it's wired into most of the pack's I2V and video-extension example workflows.

It's part of ComfyUI-LTXVideo, Lightricks' official pack (the slug reads AddLatentGuide, but the node itself is LTXVAddGuide). Keyframe conditioning is one of the things LTX genuinely does well - the 13B release made a point of it - and this is the node that exposes it.

How it works

You hand it an image, a target latent, and a frame index; it VAE-encodes the image and injects it as a conditioning anchor at that position, returning the modified conditioning and latent. The model then generates a video that passes through your anchor at the chosen frame. A strength control decides how strictly it holds to your image versus treating it as a soft hint, and the node writes into both the positive/negative conditioning and the latent so downstream samplers honor the anchor.

The inputs that matter

• positive / negative - conditioning in and (modified) out.
• vae - encodes your keyframe.
• latent - the video latent to anchor.
• image - the keyframe itself.
• frame_idx - where it lands. 0 is the first frame; negative values count from the end, which is how you pin a final frame.
• strength - how hard the anchor holds (default 1). Lower it if the keyframe makes the surrounding motion freeze.

Optionally, attention_mask confines the guide to a region, and iclora_parameters lets it participate in an IC-LoRA control setup. Outputs are the updated positive, negative, and latent - chain them into 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's under model/conditioning/ltxv. No weights of its own; the LTX checkpoint and (on LTX-2) the Gemma text encoder are the downloads.

Where people get burned

• I2V that freezes. Pin a keyframe at strength 1 and LTX can lock onto it and refuse to move - the well-known launch-era weak-I2V behavior. If your video barely animates, drop the strength and push motion through the prompt.
• Off-grid frame indices. LTX frame counts follow (8n+1) on 2.3, and keyframe positions have to make sense within that. Pinning a frame that doesn't align with the model's temporal grid gives odd results.
• Which "Add Guide" is which. The pack has several: this one (LTXVAddGuide) takes an image and adds an iclora_parameters hook; LTXV Add Guide Advanced adds CRF/blur/interpolation preprocessing controls; LTXV Add Latent Guide takes an already-encoded latent instead of an image. Pick by what you have on hand and how much preprocessing control you want.
• Skipping conditioning passthrough. This node modifies positive/negative - use its outputs downstream, not the originals, or the anchor never reaches the sampler.