LTX Director Guide Extender

This is the node you reach for the first time because of an error message, and keep because of what it does. In the LTX Director extension flow, guide images on your timeline don't apply themselves - something has to take those keyframes and bake them into the video latent and the conditioning. That something is LTXDirectorGuideExtender, the fork's reworked version of ComfyUI's LTXVAddGuide (from comfy_extras.nodes_lt). The headline fix: it appends the guide attention metadata ComfyUI expects for every inserted guide latent, which is what kills the crash ValueError: guide pre_filter_counts != keyframe grid mask length that the upstream node tripped over in extension passes.

How it works. You hand it guide_data from the timeline node (the guide_data output of LTX Director Extender, or a Prompt Relay Timeline node), plus positive/negative conditioning, the video vae, and the latent you're inserting guides into. For each guide image it encodes the image with the VAE, computes the latent-space frame index from the timeline position and the VAE's downscale formula, and appends a keyframe to the conditioning, the latent, and the noise mask at the given strength. The fork's second fix is the overlap-aware bit: extension latents often start with preserved reference frames from the previous generation pass sitting in front, and the timeline positions are relative to the new section - so the node detects that front prefix (from guide_data.frame_offset or by inspecting the noise mask) and shifts every guide insert forward so it lands in the newly generated region instead of the preserved overlap.

The inputs and outputs that matter. positive and negative are your two conditioning streams; vae is the LTX video VAE; latent is the video latent the guides get spliced into; guide_data is the timeline's guide output - connect the wrong thing here and nothing renders. scale_by (default 1.0) resizes the latent, and upscale_method (bicubic by default, but nearest-exact, bilinear, area, and bislerp are offered) picks how. Outputs are the same trio: positive, negative, and the latent with guide frames applied, ready for the sampler.

Install - same story as the rest of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/Yogurt1192/LTXDirector-Extender.git

Restart ComfyUI. No model downloads; the node leans on ComfyUI's own LTX extras, so if comfy_extras.nodes_lt is missing you're on a ComfyUI build too old for this fork (the repo targets 0.22.x).

Where people get burned. The pre_filter_counts error above is the classic one - if you still see it after installing, make sure the workflow is actually binding to the fork's node and not the upstream LTXDirectorGuide (the fork registers it under the new LTXDirectorGuideExtender id precisely so old workflows don't silently grab the upstream node). The other failure mode is a guide landing past the end of the latent - the node asserts conditioning frames exceed the length of the latent sequence, which in practice means your timeline length and your latent don't agree. That's the same manual duration-matching problem the whole pack warns about: set the duration in the extender node and the sampler subgraph to the same value.