We started with a 13B base model that already had competitive generation speeds (e.g. 55s for a 5s video on an H100 — faster than any other model out there). But we wanted to push it further to allow everyone to quickly iterate over their video generations. So we built a distilled version focused on speed without sacrificing temporal or spatial coherence.
With the Distilled model, you can now generate 5-second 720p videos in about 9.5 seconds on an H100, and around 1.5 minutes on a consumer GPU like an RTX 5090.
Because the base and distilled models are interoperable, you can mix and match them in a single rendering pipeline. That gives you three modes: - Distilled Pipeline: Fastest. Uses just 4–8 steps end to end. Ideal for rapid iteration and experimentation (~9.5 seconds on H100, ~1.5 min on RTX 5090). - Mixed Pipeline: Starts with the base model to capture accurate motion, physics, and detail, then switches to distilled at higher resolutions for faster rendering. A good middle ground. (~20s on H100, ~2.5 min on RTX 5090) - Base Pipeline: Full-fidelity generation from start to finish. Best quality, best for final renders. (~43s on H100)
All pipelines are compatible with our multiscale rendering system, which allows you to first render a lower resolution video and upscale it as needed.
The project is open source and available on GitHub (https://github.com/Lightricks/LTX-Video/) and Hugging Face (https://huggingface.co/Lightricks/LTX-Video-0.9.7-distilled). You can also enjoy our Comfy integration here (https://github.com/Lightricks/ComfyUI-LTXVideo) and our open source LTXV trainer (https://github.com/Lightricks/LTX-Video-Trainer).
Would love to hear your thoughts — especially around performance, integration into your own tools or workflows, or any creative uses you're exploring. We're actively working on tuning and adding new configurations, and early feedback is super helpful.
Creator2030•8mo ago