ComfyUI + Blender: LTX-2.3 rendering and lip sync tutorial

This tutorial builds a local AI rendering pipeline around Blender and ComfyUI, then applies it to a short commercial about a motel run by bears. The NVIDIA-sponsored presentation covers asset generation, character rigging, reference images and final video rendering with LTX-2.3.

Blender supplies depth and outline passes to guide movement and geometry. FLUX.2 [klein] supplies reference frames for appearance. For dialogue shots, a black mask over the mouth or face lets the model generate movement where strict control passes would otherwise prevent it. The speaker uses TRELLIS for character meshes and Rigify for rigging, with mesh cleanup and some manual weight painting.

Setup includes importing workflow JSON, installing missing custom nodes and downloading models into the specified folders. Output dimensions must be divisible by 64. Dialogue should contain no music and last at least as long as the video; the speaker warns that shorter audio can produce gibberish afterward.

The speaker reports about 1 minute 50 seconds for a 330-frame shot on an RTX 5090, then uses RTX Video Super Resolution for 4K output. These are reported results for this setup. Difficult shots require repeated prompting or manual reference-frame edits. Distant faces weaken lip sync, and inconsistent reference details can cause warping.