LTX-2.3 is live: rebuilt VAE, improved I2V, new vocoder, native portrait mode, and more
The AI video model, with nearly 5M downloads, tackles frozen I2V, audio artifacts, and soft details in a major update.
LTX Studio has officially launched LTX-2.3, a significant upgrade to its AI video generation platform, directly addressing widespread user feedback from nearly five million downloads of LTX-2 since January. The update tackles core issues like frozen image-to-video (I2V) outputs, audio artifacts, and soft fine details. The release introduces a native portrait mode for vertical content and a suite of architectural improvements aimed at making the model more robust and reliable for professional video creation.
The technical overhaul includes a completely rebuilt VAE (Variational Autoencoder) trained on higher-quality data, resulting in a sharper latent space that better preserves textures and edges. The team also expanded the text connector's capacity for more accurate interpretation of complex prompts and reworked I2V training to eliminate static videos. A new vocoder and audio filtering deliver cleaner sound, while the new native 1080x1920 portrait resolution caters directly to TikTok, Reels, and Shorts formats. Weights, ComfyUI workflows, and an updated API are now live, signaling LTX Studio's commitment to a production-ready pipeline for AI-generated video.
- Rebuilt VAE architecture for sharper output and better preservation of textures/edges across all resolutions.
- Native 1080x1920 portrait video generation, trained on vertical data for mobile-first platforms like TikTok and Reels.
- Reworked I2V training to reduce freezing and improve motion, plus a new vocoder for cleaner audio with fewer artifacts.
Why It Matters
This makes AI video generation more reliable and practical for creators, directly addressing the pain points that hindered professional use.