LTX 2.5 in ComfyUI: setup, LoRAs and low-VRAM options

This tutorial covers LTX 2.5 setup in ComfyUI for offline video generation. It assumes ComfyUI is already installed and starts with an update, then explains which templates and model files to use. The speaker selects the local workflows rather than the paid cloud templates.

The walkthrough covers text-to-video, image-to-video and first/last-frame generation. Required downloads include a diffusion model, spatial upscaler, text encoder and audio and video VAEs. The speaker recommends the distilled model for generation, describing four to six steps versus roughly 20 to 30 for the dev model. Text and image workflows generate a low-resolution pass before upscaling; the first/last-frame template shown lacks that upscaler.

For hardware planning, the speaker cites a 16 GB VRAM minimum. A community Q3 small GGUF file is 12.6 GB, which they suggest may work with 12 GB VRAM. That is a possibility, not a demonstrated hardware guarantee. The tutorial also shows how to connect an existing LTX 2 LoRA and add its trigger word.

The speaker reports roughly 20-second generation times for two examples and faster results than MiniMax H3 on their computer. These are example timings, not general benchmarks. The closing section introduces LTX Director and warns that using the LTX upscaler on other models can give uneven results with high-action footage.