Descriptions:
Nerdy Rodent walks through two practical ComfyUI workflow techniques built around LTX Video 2.3, the open-weight video generation model. The first segment addresses a common frustration: CLIP’s built-in content filtering causing prompt refusals even for innocuous requests. The video demonstrates the abliterated CLIP LoRA — a modified encoder that removes the safety filtering layer — showing side-by-side outputs where a refused prompt like “the duck does something very harmful” generates an actual video instead of an apology loop. As an alternative, the workflow also integrates Ollama with Gemma 3 4B as a vision-capable local prompt enhancer that sidesteps the issue entirely.
The second half covers the Union Control LoRA for LTX 2.3, which consolidates canny edge, depth, and pose control into a single model rather than requiring separate LoRAs for each. The tutorial uses the LTX IC LoRA loader specifically because of its latent downscale factor output, which is required for the Union model to function correctly. Pose generation is handled by the SD pose model, noted as faster and more accurate than DW pose in the author’s testing.
All workflows follow the “Rodent Method” — color-coded node groups designed for readability and easy modification. The video includes downloadable ComfyUI workflows via Patreon and is aimed at intermediate ComfyUI users already familiar with LTX basics who want finer control over video generation with open-source tools running locally.
📺 Source: Nerdy Rodent · Published March 13, 2026
🏷️ Format: Tutorial Demo
