LTX 2.3

By Lightricks · Updated

What It Actually Is

LTX 2.3 is Lightricks’ open-source chess move against every closed video API: a 22-billion-parameter diffusion transformer that generates cinema-quality video and synchronized audio from a single text or image prompt — and it runs on your own hardware. No subscription. No credits. No content filters deciding what you can and can’t create. The model produces up to 20-second clips at native 1080p with built-in sound effects, dialogue, and ambient audio in one forward pass. Need 4K? A two-stage upscaling pipeline gets you there. The entire stack is open-weight under Apache 2.0, with day-one ComfyUI support and a standalone desktop app. For the first time, a local video model can genuinely compete with the cloud giants.

Key Strengths

• Native audio-video sync: Generates perfectly matched sound — dialogue, ambient noise, effects — in one forward pass. No separate pipelines, no post-processing.
• Fully open-source (Apache 2.0): Full commercial use, local deployment, and fine-tuning allowed. Your GPU, your rules.
• Redesigned VAE: Sharper textures, cleaner edges, and realistic fine details compared to prior open-source video models.
• Fast distilled variant: The 8-step distilled model runs blazingly fast on consumer GPUs (RTX 3090+), making real-time iteration viable.
• Native portrait mode: 1080×1920 output ready for mobile-first and social media content without cropping or letterboxing.

Honest Limitations

• Trails closed leaders on fidelity: Kling 3.0 and Seedance 2.0 still produce more polished, physically accurate output in complex scenes.
• Audio artifacts: Occasional background music bleed, slow-motion drift in longer clips, and minor visual glitches in demanding scenes.
• Multi-shot timing: Complex multi-scene prompts can produce uneven pacing; iterative generation plus stitching gives best results.
• 4K is VRAM-heavy: The two-stage upscaling pipeline demands serious GPU memory. Single-stage outputs cap at 1080p.