LTX-2.5 first generates video and audio at half resolution, then upscales and refines them; this LoRA is used only in the first stage, reducing the distilled version from 8 steps to 2 steps, applicable to both text-to-video and image-to-video generation. The author trained it using DMD2 on a single H200 GPU; it is currently a candidate version at step 12000, with plans to train up to step 24000. Training resolutions are 576×320 and 320×576, with refined output at 1152×640 and 5 seconds. The accompanying ComfyUI workflows also incorporate NVIDIA's one-step refinement model and Pruna's streamlined VAE decoder, with separate setups for local use and Comfy Cloud.
Why it matters With one-step refinement and a faster decoder, short video previews with audio can be generated more quickly, making it suitable for initial prompt and seed selection.
Hugging Face Likes 5 · Hugging Face Trending Score 4 · Hugging Face Downloads 27
License: Conditional Commercial Use · ComfyUI: Mentioned in Model Card · VRAM: Not Specified
Reasons
- The model card provides complete sampling parameters (sigmas, sampler, CFG) and multiple examples for text-to-video and image-to-video generation
Usage
- Download the corresponding t2v or i2v workflow; apply the LoRA only in the first stage with strength 1.0, 2 steps, Euler Ancestral, and CFG 1; do not apply it to the refinement stage
Risks
- The author states it is experimental, suitable for quick previews but not for final production
- Under the LTX-2.x community license, commercial use is conditional; other resolutions have not been trained