Lightricks 2026.03 Release

LTX-2.3

The new benchmark for open-source AI video. Single-model DiT architecture with native audio sync.

20s
Max Duration
1080P
Native Vertical
50 FPS
High Framerate
Sync
Audio Sync

What is LTX-2.3?

LTX-2.3 is an open-source AI video generation model by Lightricks, representing the latest iteration of the LTX-2 family. It is a single-model Diffusion Transformer (DiT) that generates high-fidelity video and synced audio simultaneously.

Supporting T2V, I2V, and A2V modes, it is hailed as the "open-source Veo 3", offering zero-cost local runs, fast speeds, and quality rivaling top closed-source models.

LTX-2.3 vs LTX-2

FeatureLTX-2 (Old)LTX-2.3 (Current)
VAE & Latent SpaceStandard resolution✅ Rebuilt for 40% sharper textures
Prompt AdherenceStruggles with complex instructions✅ 4x text capacity with gated attention
I2V ConsistencyOccasional "frozen frames"✅ Ultra-consistent, fewer artifacts
Audio QualityNoticeable background noise✅ Studio-clean with MS-level sync

LTX-2.3 Advanced Prompting

Chronological order: Describe sequences step-by-step.

Cinematic keywords: Use "Close-up", "Crane shot", etc.

Lighting: Add "Volumetric light", "Neon", etc.

Audio tags: Mention "Heavy bass", "Rain sounds" in prompts.

Core Improvements

Optimized for Professional Production

Sharper Details

Clearer hair, textures, and edges.

Stronger Following

New gated attention for complex prompts.

Realistic I2V

Reduced frozen frames and better consistency.

Cleaner Audio

Optimized filtering + new vocoder.

Highlights

Native Vertical

1080x1920 for TikTok/Shorts.

Audio-Guided

Audio drives motion and lipsync.

Multi-modal

Supports Keyframes, Depth/Pose/Canny.

Specifications

Resolution Up to 1080p (Native Vertical), 4K Upscale
Duration Up to 20s (extendable)
Framerate 24 / 48 / 50 FPS
Model Size Approx. 22B Parameters
Speed Seconds locally
License LTX Community (<$10M Free)

Download

Available on Hugging Face.

LTX-2.3-22B-Dev

Full bf16 version, best quality.

LTX-2.3-22B-Distilled

Distilled version, 8 steps efficiency.

FP8 Quantized

Best for 12GB VRAM.

Run Locally

  • 01 Recommended: ComfyUI-LTXVideo.
  • 02 Supports official scripts, CLI, and Fal.ai.
  • 03 LTX Desktop: Open-source professional editor.

Notes

Use chronological prompt structure. Quantization recommended for low VRAM. Follow community safety guidelines.

"LTX-2.3 is the ultimate open-source solution for synced video/audio."

© 2026 wan2.video