LTX-2.3 Crop Video Guide ⚡

This is the least glamorous node in the LTX-2.3 section and the one people forget until their videos come out wrong. LTX-2.3's video-to-video mechanism doesn't partially denoise your source - it VAE-encodes the source's frames and appends them as extra guide tokens the model cross-attends to during sampling. Those tokens are still sitting in the latent when sampling finishes. If you decode straight away, you get the guide frames glued to the front of your output. This node strips them back off.

It's a thin, correct wrapper around comfy-core's own LTXVCropGuides (the keyframe_idxs RoPE machinery). The reason it exists as a separate node is that core's version only understands a plain video latent, while this pack's LTX-2.3 joint AV latent is a NestedTensor((video, audio)) pair - so the node splits the streams, crops the video half, and rejoins them. When no guide was ever appended (e.g. you used ic_lora="none" upstream), it's a no-op that returns the latent untouched. You can leave it in the graph permanently and it costs you nothing.

What it takes

Three required inputs, all pass-through:

• positive (CONDITIONING) and negative (CONDITIONING)
• latent (LATENT)

Outputs are the same three, cleaned. The canonical LTX-2.3 vid2vid chain is: LTX-2.3 VidToVideo → LTX-2.3 KSampler (distilled) → this node → LTX-2.3 AV Decode. It also sits in the EditAnything and person-replacement recipes, always in the same spot: after the sampler, before decode - and before any latent upscale, because the crop needs to happen while the guide coordinates are still aligned.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/ChrisColeTech/ComfyUI-GGUF-Loader
pip install --upgrade gguf

Restart and it's under 🤖 CCTech/LTX-2.3. No model files, no extra dependencies - it's pure tensor surgery.

Gotchas

Don't confuse this with cropping the video to a resolution. It crops nothing you can see; it removes hidden reference frames from the latent. And if you're hand-building a graph, remember the ordering constraint: crop before upscaling (via LTX-2.3 Latent Upscale x2) or before decoding. Upscale first and the keyframe coordinates have moved under you.