Lightricks just released LTX-2.3. It dropped last night. Open-source, free to run locally, and it might be the most complete AI video model update we have seen in months.
What is LTX-2.3?
LTX is Lightricks's open-source AI video generation model. You can run it locally via Hugging Face or in ComfyUI, which added day-0 support today. No subscription. No server fees. Your footage stays on your machine.
LTX-2.3 is a 20B parameter audio-video model. That means it generates video and audio together from a single prompt. Version 2.3 brings six major upgrades:
- Native 9:16 portrait video — generates vertical format directly, not cropped from landscape. Cleaner quality for Reels and TikTok.
- Sharper details — a redesigned VAE architecture delivers crisper textures and cleaner edges in every frame.
- Better audio — cleaner sound, improved dialogue, music, and ambient layers.
- Smoother image-to-video — fewer frozen frames and glitches, more natural motion from still images.
- Smarter prompts — the updated text encoder follows complex instructions much more accurately.
- Text rendering — it can render readable text inside video frames now. This was broken in most models.
Why It Matters
Open-source AI video has been trailing commercial tools like Runway and Kling for a while. LTX-2.3 closes that gap in ways that actually matter for production work.
The portrait mode is the headline feature. Most AI video tools force you to generate 16:9 then crop or letterbox to 9:16. That kills resolution and quality. LTX-2.3 generates 9:16 natively. For social content production, that alone is significant.
The local-first approach matters too. When you are working with client footage or brand assets, you do not always want that material passing through third-party servers. Running LTX-2.3 on your own hardware means your client data stays private.
And with ComfyUI day-0 support, the node-based workflow ecosystem already has templates ready. You can pull the T2V or I2V workflow templates straight from the template library and start generating today.
HOW SEQNCE WILL USE THIS
We are evaluating LTX-2.3 closely. The native portrait generation and local-first privacy model make it interesting for client work in regulated industries. The audio-video generation is something we want to test for atmospheric B-roll with synchronized ambient sound.
Open-source tools like this are on our radar. When the quality clears the bar for commercial deliverables, they go into the stack. We will report back once we have run proper tests.
Quick Takeaways
- Native 9:16 portrait video, no quality loss from cropping
- Simultaneous audio and video generation in one model
- Fully open-source, free, runs locally on your own hardware