LTX Director Guide

LTX Director is the pretty face of this workflow, but LTX Director Guide is the engine room. It's the node that takes the timeline's payload - the guide_data and motion_guide_data from LTX Director - and does the actual heavy lifting: VAE-encoding your keyframe images, inserting them into the video latent, building the noise mask that freezes or frees frames, and applying IC-LoRA motion guidance if you've got a control video on the timeline. If you load a Director workflow and see a big node with a lot of inputs sitting between Director and the sampler, this is it.

How it works

Split it into two paths. The standard path is the classic LTX guide mechanism: for every image segment in guide_data it encodes the image with the video VAE, writes it into the latent at its assigned frame (pixel-space positions converted through the VAE's scale factors), marks those frames in the noise mask, and appends the keyframe info to your positive and negative conditioning so the model knows what each anchored frame is.

The motion path is where 2.0's IC-LoRA support lives. Connect motion_guide_data (videos you dropped on the Director motion track), a model, and pick an IC-LoRA from ic_lora_name - those files live in ComfyUI/models/loras, and if the dropdown shows only put_ic_lora_in_ComfyUI_models_loras, that's a literal instruction, not a placeholder bug. The node loads the LoRA onto the model, downscales the latent to match the LoRA's expected grid (latent_downscale_factor tells you what it did), and uses the motion video segments to condition movement. The author chose the LTX Ingredients IC-LoRA route over MSR/BFS after testing, because he found it the most stable.

There's also a Retake Mode branch: instead of inserting stills, it loads your base video, encodes it, and uses a temporal noise mask so only the selected segment regenerates while the rest stays frozen - the "redo this shot" feature. Retake auto-detects from the timeline unless you force it with retake_mode.

The inputs that matter

• positive / negative / vae / latent - your conditioning pair, the video VAE, and the latent to write into.
• guide_data - required, straight from LTX Director.
• ic_lora_name / ic_lora_strength / model - only if you're using motion guidance. Strength defaults to 1.0; drop it if motion feels overpowered.
• scale_by / upscale_method - spatially rescale the latent (handy when your IC-LoRA expects a smaller grid than your output).
• use_tiled_encode / tile_size / tile_overlap - tiled VAE encoding for tight VRAM. Turn it on if the plain encode OOMs.
• retake_mode - force Retake Mode on/off.

Outputs

positive, negative, latent (the latent with guides inserted and noise mask set), model (with the IC-LoRA patched in - feed this to the sampler, not the Director's model, when using motion guidance), and latent_downscale_factor (FLOAT) so downstream nodes know how the latent was scaled. After this node, run LTX Director Crop Guides to trim any padding frames, then sample.

Install

It ships with the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/WhatDreamsCost/WhatDreamsCost-ComfyUI

Or search WhatDreamsCost in ComfyUI Manager. No extra Python deps, but the README's rule applies: update ComfyUI-LTXVideo and ComfyUI-KJNodes to current before running Director workflows, and IC-LoRAs go in ComfyUI/models/loras.

Gotchas

Retake Mode will throw a clear error if no base video is on the timeline - the message tells you to wait for the upload overlay, so if you see it, the upload genuinely didn't finish. Watch the latent_downscale_factor if you're hand-wiring a custom workflow: a sampler fed the original latent dimensions after the Guide downscaled for an IC-LoRA will produce garbage-shaped output. And if motion segments look like they're being ignored, check that both motion_guide_data and a model are connected - the IC-LoRA branch only runs when both are present.