Muse Crop Guides

If you've used an IC-LoRA or image-reference guide with LTX, you've seen the failure mode: the draft looks right, then the Stage 2 upscale pass crops or resizes the latent, and suddenly the guide frames are the wrong shape and the conditioning drifts. MuseCropGuides exists to fix exactly that - it trims your guide keyframes to match the cropped or upscaled latent so the reference conditioning lines up with what the sampler is actually working on.

This is the kind of node that sounds like a non-feature until the day it saves a whole workflow. It's a small, focused fix, and it comes from a real problem: LTX Director's two-stage flow generates at half resolution, then the spatial upscaler refines at full res, and any crop the upscaler applies silently desyncs your reference frames from the latent they're supposed to guide.

What it does

Three sockets in, three sockets out. It takes positive and negative conditioning plus a latent, and returns the same three, with the guide conditioning adjusted to the latent's current geometry. You drop it in the graph between the guide encoder and the sampler - in a Muse workflow, that's typically after the Stage 2 upscaler, on the path that feeds the refinement pass.

It's the companion to MuseGuide (which does the encoding in the first place) and it ships in the same pack. In the original WhatDreamsCost codebase these two were part of one monolithic node; Muse reimplemented them as standalone pieces precisely so you could drop them into a vanilla LTX graph without pulling in the whole director.

Why it matters

The spatial upscaler changes the latent's dimensions. Guide keyframes were encoded to match the pre-upscale shape. Feed stale conditioning into the refinement pass and you get reference frames that are cropped wrong, stretched wrong, or just plain misaligned - which reads as style bleed, ghosting, or the model ignoring the reference entirely. Cropping the guide to the new latent geometry keeps the IC-LoRA reference honest.

The honest caveat: this is a plumbing node. There's nothing to configure, no sliders to tune, and if you're not running a two-stage guide pipeline you don't need it. But if you are, it's the difference between a reference that holds through the upscale and one that quietly falls apart.

Installing it

Part of the muse-ltx-timeline pack:

cd ComfyUI/custom_nodes
git clone https://github.com/muse-collective-26/muse-ltx-timeline

Restart ComfyUI. No extra Python deps beyond the pack's usual trio (pip install av torchaudio soundfile), and no model downloads of its own - it works on whatever LTX 2.3 setup you already have.

Where it sits

For the record, the input/output names it uses - positive, negative, latent - are standard ComfyUI conditioning types, so it'll wire into any graph that has those sockets, not just Muse's. If you've built a manual LTX 2.3 workflow with a guide encoder and a separate upscale pass, you can graft this node in to fix the same mismatch. That's the nice thing about Muse's decision to keep the guide utilities standalone: they play well with other people's graphs.

Use it once and it just quietly does its job. Which is, for a node that exists to prevent a specific misalignment, exactly the right level of excitement.