LTX 2.3 in ComfyUI: inpainting and interpolation tutorial

Learn to edit LTX 2.3 videos with LoRAs and SAM 3 masks, preserve source audio, and use FILM to interpolate a 24 FPS clip to 48 FPS.

Player not loading? Watch on YouTube

This tutorial extends a ComfyUI workflow for LTX 2.3 with Adaptive Projected Guidance (APG), community LoRAs and masked video edits. The presenter compares generations with APG enabled and disabled, judging the enabled result closer to the cinematic prompt. He says APG has a more visible effect when CFG is not 1, though it also works with distilled models.

The video editing examples explain how stage-two processing, empty latents and denoise settings affect changes to an input clip. One example uses a linear quadratic scheduler with denoise set to 0.42. Empty latents also allow an aspect ratio change.

For inpainting, SAM 3 selects areas such as a face, window or clothing. The presenter notes a 32-token segmentation prompt limit. Guide nodes and mask colors depend on the LoRA: one requires the IC LoRA guide and black masks, while another accepts the usual guide node and white masks. Expanding a mask can help avoid retaining the original object's shape. A clothing edit keeps the input audio by disabling generated audio in the output settings.

The final example uses FILM through ComfyUI's native frame interpolation node. A multiplier of two converts a 1280 by 704 clip from 24 FPS to 48 FPS, with interpolation available for the generated clip or its extended version.