LTX Video 2.3

LTX Video 2.3 (LTX-2.3) is the latest open-source video generation model from Lightricks, and one of the most capable video models available to self-host. Built on a DiT (Diffusion Transformer) architecture with 22 billion parameters, it sets a new bar for open-source video quality.

**What makes LTX 2.3 stand out**: - **True 4K resolution** — Generates native 4K video, not upscaled 1080p. Sharpness and fine detail are handled by a newly designed VAE component. - **50 FPS** — Smooth, cinematic motion at up to 50 frames per second - **Native synchronized audio** — Audio is generated in the same pass as video, not added in post. LTX 2.3 is one of the first open-source models to natively synchronize audio and video. - **20-second clips** — Longer than most open-source competitors, enabling more narrative content - **Apache 2.0 license** — Fully open for commercial use; weights, training code, and inference pipeline are all public

**Architecture highlights**: LTX 2.3 uses a new VAE specifically designed for fine visual detail, combined with a DiT backbone optimized for temporal consistency. Text-to-video, image-to-video, and video-to-video modes are all supported.

**Access**: Available via the official GitHub repository, Hugging Face weights, and inference APIs on platforms like fal.ai and BestPhoto. The model can be self-hosted on high-VRAM GPUs (40GB+ recommended for 4K inference).

LTX 2.3 makes cinema-quality AI video generation available to researchers, indie creators, and studios without proprietary lock-in.