✂ LTX Likeness Crop

This is the simplest node in the pack, and the one most people can safely ignore - which is exactly why it's worth a straight answer about what it does. LTXLikenessCrop takes a LATENT, crops the appended likeness reference frames off the end, and returns a LATENT of the original length. It's the inverse of LTXLikenessGuide's emit_latent=extended mode: the Guide appends a reference frame to the latent, and this node removes it before VAE decode so you don't decode a reference frame into your output video.

One input (latent), one optional reference_info (used to work out exactly how many frames to crop when the metadata is available), one output (LATENT), and a debug flag.

When you need it - and when you absolutely don't

The crucial detail, straight from the Guide's own documentation: with the Guide in its default passthrough mode, this node is unnecessary. The default emit_latent=passthrough means the Guide's latent output is the unmodified input latent - no extension, nothing to crop. Wire the latent straight to VAE Decode and forget this node exists.

You only need LTXLikenessCrop if you explicitly set the Guide to emit_latent=extended, which is the legacy mode that appends the reference frame into the latent itself. In that mode the latent is longer than your video, the reference frame is riding along at the end, and you need this node to strip it before decode. There's a good reason the default changed: the extended mode can trigger end-keyframe interpolation pressure even in silent_reference mode, because the spatial pattern itself of an appended reference can invoke the model's learned keyframe behavior. The Guide's real identity-preservation value lives in its conditioning modifications and reference_info metadata, not in latent surgery - which is why the default is now passthrough.

So the practical guidance is: if you inherited a workflow with LTXLikenessGuide → ... → LTXLikenessCrop → VAEDecode, check the Guide's emit_latent setting. If it's passthrough, the Crop node is a harmless no-op that preserves validation (its reference_info is optional and it degrades gracefully). If it's extended, the Crop is load-bearing - removing it will decode an extra reference frame into your video. It's a small, safe, and occasionally essential piece of plumbing, not a tuning node.

Install is the pack-wide clone (10S_Nodes into custom_nodes, restart, or ComfyUI Manager → "10S-Comfy-nodes"), no extra deps, LTX2-class models only.