LTX-2.3

An open-weight diffusion-transformer model that generates synchronized video and audio from text, image, and audio conditioning.

LTX-2.3 is Lightricks' open-weight diffusion-transformer release for synchronized video and audio generation. Instead of predicting the next text token, it denoises joint audio-video latents so motion and sound stay aligned in the finished clip.

At a glance

Family
ltx
Source type
Open weights
Released
March 2026
Authors
Lightricks
Modalities
Text, Image, Audio, Video, Multimodal
Parameter count
22 billion parameters

What It Is

LTX-2.3 is an open-weight video and audio generation model from Lightricks, released on 2026-03-05 with about 22 billion parameters. Public materials describe it as a diffusion transformer built for joint audio-visual generation rather than text-only language modeling. The release ships as downloadable weights on Hugging Face and is positioned as the latest member of the LTX video model line.

Inputs And Outputs

LTX-2.3 accepts text prompts and can also use image or audio references as conditioning signals. Its primary outputs are synchronized video with matching audio rather than a text reply or a still image alone. That multimodal shape matters because viewers judge clips by both picture and sound, so the model is framed around joint generation instead of stitching video and audio together after the fact.

Architecture

At a high level, LTX-2.3 follows a latent diffusion pattern. Encoders compress prompts and optional reference media into conditioning features, a diffusion transformer denoises corrupted audio-video latents over many steps, and decoders render pixels and waveforms. Cross-attention lets the denoiser read text and reference signals while it updates the latent sequence. This is a different job from an autoregressive text model: the network learns to remove noise from a compressed spatiotemporal representation instead of emitting one vocabulary token at a time.

Important Modules

Cross-attention is the main named module on this page because it carries conditioning from text and reference inputs into the denoising transformer. That fusion step is what lets a prompt or reference image steer motion, framing, and audio tone while the latent sequence is still being refined.

Training

Public LTX materials frame training around a diffusion objective: the model learns to predict denoising updates that recover clean audio-video latents from noised versions. That objective matches the generation loop at inference time and differs from next-token cross-entropy training used by text-only language models.

Training regimes

Linked papers

Practical Notes

LTX-2.3 is useful to read as a latent video diffusion system with synchronized audio rather than as a chat model with a video export button. Working in latent space keeps generation tractable, but readers still need encoder and decoder stages to move between user-visible media and the tensors the transformer denoises. The open-weight release makes the architecture inspectable, while the LTX-2 paper and model card remain the best places to confirm implementation-specific details.

Tags

References

  1. Lightricks. "Lightricks/LTX-2.3." Hugging Face, 2026.
  2. Lightricks. "LTX-2.3: Introducing LTX's Latest AI Video Model." LTX, 2026.
  3. Lightricks LTX Team. "LTX-2: Efficient Joint Audio-Visual Foundation Model." arXiv, 2026, https://arxiv.org/abs/2601.03233.
  4. Lightricks. "Lightricks/LTX-2." GitHub, 2026, https://github.com/Lightricks/LTX-2.