LTX 2.3 in ComfyUI: model comparison and video extension

This tutorial covers LTX 2.3 version 1.1 in ComfyUI, with a focus on extending generated clips and existing videos. The presenter compares Kijai's FP8 conversions with the official BF16 checkpoint and a GGUF UDQ4 option. The model update promises better audio and a different aesthetic; it also has an updated Union Control LoRA for version 1.1 compatibility.

The comparison uses 24 FPS, 121 frames at 1280 by 704, eight steps and a fixed seed. In the presenter's tests, FP8 and BF16 take similar amounts of time, while GGUF takes several minutes longer overall. Those timings describe this setup rather than a general performance guarantee. The presenter finds BF16 faces slightly more detailed and the GGUF sample's audio somewhat weaker.

For local AI video editing, the workflow uses an LTXV audio video mask node to select where an extension begins and ends. A separate prompt can continue a scene or change its direction. Examples replace dialogue near the end of an existing clip, extend a 5.8-second clip to 11 seconds, and create alternate endings by changing the prompt. A final example starts with an image. The presenter judges the updated audio slightly better but notes that unwanted background music remains a problem unless the prompt specifies an ambiance.