LTX-2.3 in ComfyUI: BF16, FP8 and workflow tests

This tutorial tests LTX-2.3 in ComfyUI for local AI video generation with audio. The presenter compares the 22B dev model, its FP8 version and Kijai's split files using the same difficult transition between a starting image and an ending image. The initial test uses 121 frames at 24 FPS and 1280 by 720 resolution. Results vary: some clips follow the woman's movement better, while others handle the dolphin-to-orca transition more convincingly. The presenter reports lower VRAM use with FP8 and describes the model's license as non-free.

Further tests remove the image inputs, swap the 2x upscaler for a 1.5x version and rewrite prompts with the LTX2 AI prompter or Ollama running Qwen 3 VL. Small prompt changes produce noticeably different clips, but neither rewriting approach consistently fixes the requested actions or spelling.

For 1920 by 1080 generation, the presenter uses a chunk feed forward node to reduce VRAM use and suggests increasing the chunk count after an out-of-memory error. A distilled example uses eight steps at CFG 1, followed by a longer 241-frame test. Visible limitations include malformed fingers, incorrect text and slight ghosting that the presenter attributes to doubling the FPS. The tutorial also demonstrates custom audio and a 720 by 1280 portrait clip, then points to ComfyUI's built-in example workflows.