LTX-2.3

LTX-2.3 is an image-to-video generation model that produces synchronized video and audio from an input image and text prompt. Built on a diffusion-based architecture, it is designed for high-fidelity audiovisual content creation and can be accessed as a managed, OpenAI-compatible API through gigarouter.

The model uses an asymmetric dual-stream transformer with a 14-billion-parameter video stream and a 5-billion-parameter audio stream. Bidirectional cross-attention layers with temporal positional embeddings and cross-modality AdaLN enable tight audiovisual alignment. A multilingual text encoder broadens prompt understanding, and modality-aware classifier-free guidance (modality-CFG) improves controllability of both video and audio outputs. The audio stream can generate speech, background sounds, and foley consistent with the scene.

According to the LTX-2 paper (arXiv:2601.03233), the model achieves state-of-the-art audiovisual quality and prompt adherence among open-source systems, while delivering results comparable to proprietary models at a fraction of the computational cost and inference time. A distilled variant (ltx-2.3-22b-distilled) runs in 8 sampling steps at CFG=1 for fastest inference.

For production use, the recommended pipeline is the two-stage text/image-to-video pipeline (TI2VidTwoStagesPipeline), with a higher-quality variant using a different sampler. Additional supported modes include video-to-video, keyframe interpolation, retake, IC-LoRA, lipdub, and text-to-audio.