Lightricks drops LTX-2.3 with native vertical video and a free local editor
Lightricks just launched LTX-2.3, upgrading its open-source audiovisual model with sharper textures, cleaner dialogue, and native 1080x1920 portrait video. Alongside it comes LTX Desktop, a free local video editor powered by the same engine. Why it matters: Full model weights and ComfyUI workflows are live on Hugging Face, free for companies making under $10M in annual revenue.

LTX-2.3 is an open-source multimodal engine from Lightricks that generates synchronized video and audio in a single pass, accompanied by a free, local desktop editor called LTX Desktop.
Instead of treating video and sound as separate problems, the model is built to handle both simultaneously while fixing the most annoying bottlenecks of earlier checkpoints.
What it can do
The biggest visual change is native vertical video support up to 1080x1920. It is trained directly on vertical data rather than cropped from landscape footage, making it instantly usable for mobile-first platforms like TikTok, Reels, and YouTube Shorts.
Under the hood, Lightricks rebuilt its VAE architecture to create a cleaner latent space. Fine textures like hair and object edges are noticeably sharper across resolutions, reducing the need for post-generation upscaling.
Image-to-video reliability also got a major overhaul. The team tuned out the static freezing and slow pan effects that often ruined I2V outputs, giving the engine more dynamic, realistic motion while maintaining input frame fidelity.
Audio output sounds far more natural. Thanks to a new vocoder and a filtered dataset that stripped out noise and silence artifacts, generated dialogue is clean and closely aligned with character movements.
Under the hood and how to run it
LTX-2.3 is a 22B-parameter DiT foundation model that pairs with Google's Gemma 3 12B text encoder for prompt interpretation. Its expanded text connector significantly improves spatial tracking and multi-subject adherence.
Lightricks open-sourced the entire stack on Hugging Face: the base 22B weights, an 8-step distilled checkpoint, distilled LoRAs, and latent upscalers. Official ComfyUI reference workflows are available right out of the gate.
Along with the model comes LTX Desktop Beta. It is a completely local, open-source video editor built directly on top of LTX-2.3. Once installed, it runs entirely on your own GPU without requiring an internet connection or credit tokens.
Lightricks kept the licensing developer-friendly: the weights and desktop app are free to use and commercialize for any entity making under $10M in annual revenue.
Why it matters
We think LTX-2.3 is a massive win for creators who want production pipelines without cloud fees or subscription lock-in. Shipping native portrait aspect ratios and clear dialogue out of the box eliminates hours of post-processing.
Running a 22B model alongside Gemma 3 locally will demand serious hardware, but the distilled 8-step weights show Lightricks is committed to practical local execution. It sets a high bar for open-source media engines.
Sources (2)
- ltx.io(opens in a new tab)Originalltx.io
- huggingface.co(opens in a new tab)Reporthuggingface.co