LTXV Multi Guide (N Frames)

LTXV Multi Guide (N Frames) is how you tell LTX Video "this scene starts here, and this other image shows where it should end." It takes a batch of reference images and injects each one into the latent at a specific point in the timeline - first frame/last frame keyframing, shot sequences, multi-image storyboarding. LTX is the speed-first video model that drafts clips in seconds on consumer GPUs (the KB tracks it at ~20s per 2B clip on a 4090), and its weak spot is control; this node is a control tool, letting you pin any number of frames instead of just one.

It's a faithful mirror of WhatDreamsCost's LTX Sequencer - a well-known community node pack - implemented so the job works with a dynamic N instead of a fixed chain. Mechanically it wraps LTX's native LTXVAddGuide.append_keyframe in a loop: each reference image is VAE-encoded, its target frame is converted from pixel frames to latent frames using the VAE's downscale formula, and the keyframe is appended to the conditioning. It deliberately skips the manual attention-entry trickery and returns a clean latent dict, which is the difference between "works reliably across ComfyUI updates" and "breaks whenever the internals shift."

The inputs that matter, in order. positive / negative / vae / latent are the standard LTX stack from your sampler and VAE - wire these through, since the node returns an updated version of each. images is your reference batch. mode (frames, seconds, or percentage) decides how you express where each image goes, and fps (default 24) converts seconds/percentage into frame numbers. strength (0–1, default 1) is the global guide weight, and it's overridden per-image if you fill in strengths as a comma-separated list.

Then the three ways to position images: connect indices (an INT list - e.g. from the pack's Image List Sampler) and it wins outright; otherwise type comma-separated positions in positions (in whatever mode you picked, and -1 means last frame in frames mode); if neither is given, images are distributed evenly across the clip. The outputs are positive, negative, and latent - the updated stack - which loop straight back into your sampler. That's the mental model: it sits between the VAE/latent setup and the KSampler, like a filter that conditions the whole clip on your references.

Gotchas to respect. It needs LTX support in ComfyUI core (comfy_extras.nodes_lt) - that ships with modern ComfyUI, but if the node errors on import, update ComfyUI first, not the pack. The images input is list-aware and flattens batches, so don't worry about list vs batch wiring. And a mismatch between image count and position count is handled by truncating to the smaller with a console warning - silently losing guides if you mistyped a list, so keep them in sync.

Install: ComfyUI Manager → search "AnotherUtils", or cd ComfyUI/custom_nodes && git clone https://github.com/marcoc2/ComfyUI-AnotherUtils, then restart. No extra Python deps - it leans on ComfyUI's LTX nodes, not on Lightricks' separate node pack.