LTX 2.3 in ComfyUI: Union control and CLIP LoRA tutorial

This ComfyUI tutorial covers LTX 2.3 video workflows with Union control and an abliterated CLIP LoRA. The presenter first compares model variants and finds their example outputs broadly similar. He says scaled models should run slightly faster on Nvidia 4000 series cards or later.

The CLIP LoRA section addresses refusals from the LTX2 prompt enhancer. A duck prompt produces a refusal that carries into the generated video's speech; loading the LoRA produces a different enhanced prompt. The presenter says this LoRA does not change the video model and ideally should not handle encoding. He also warns that the prompt node's default length of 256 can truncate longer prompts. Gemma 3 4B provides another prompt-enhancement example.

Union supports Canny, depth and pose guidance in one model. The workflow uses the LTX IC LoRA loader for its latent downscale factor output, which the presenter says is necessary for correct scaling. Demonstrations focus on pose and Canny rather than depth.

Results vary: simple head movements follow the pose input, while a gesturing robot has awkward hands. A retry at 1920 by 1088 with thicker pose lines appears better to the presenter, though requested chalk artwork remains absent. A spinning-person test exposes motion problems, while Canny-guided animated lettering fares better. The final example still misses some requested colors and direction details.