New VAE architecture
The rebuilt autoencoder is the headline change in 2.3: sharper textures, cleaner faces and small objects that survive the trip to 4K rather than smearing.
Lightricks
LTX-2.3
Open-source 4K video, up to 20 seconds. An Apache 2.0 video model that renders up to 20 seconds at 4K with native audio, a rebuilt VAE for sharper detail, and endpoints for audio-driven, extended and re-taken shots.
LTX-2.3 is not in the Studio catalog yet — the buttons open our text-to-video and image-to-video workflows, where comparable models are available today.
- Max duration
- 20s
- Max resolution
- 4K (2160p)
- Audio
- Native, plus audio-to-video
- Frame rate
- 24 or 48 fps
- Aspect ratios
- 16:9, 9:16
Features
What LTX-2.3 does differently.
Native audio and a dedicated audio-to-video endpoint
Audio is generated in the same pass and is noticeably cleaner in 2.3. A separate audio-to-video endpoint goes the other way: feed a track and let the picture follow the sound.
Extend and retake
extend-video adds another 2 to 20 seconds onto a clip you already have, and retake-video regenerates a shot without restarting the project.
Open weights under Apache 2.0
The weights are published, so you can run LTX-2.3 locally and LoRA fine-tune it on your own characters or house style, then keep using the hosted endpoints for scale.
Portrait, high frame rate, last-frame interpolation
Native 9:16 for mobile, a choice of 24 or 48 fps, and last-frame interpolation so a shot ends where the next one needs to begin.
Examples
Prompts that show its range.
Snorkel lens ground-scraping tracking shot barefoot Ethiopian runner dawn, feet kicking red dust 240fps slow-mo, Rift Valley, Lubezki natural light
Try this promptDrone descent through oculus of derelict Soviet radio telescope, botanist cataloging wildflowers, red jacket only color, Tarkovsky Stalker
Try this promptThrough-the-veil bride's face Indian wedding, sheer red dupatta textured overlay, 85mm f/1.2, warm tungsten candlelight, Mira Nair intimacy
Try this promptA woman whispering into the microphone
Try this promptLTX-2.3 renders each example on demand rather than hosting fixed sample clips, so the prompts are reproduced here instead. Run one to see your own render.
Workflows
Every way to run LTX-2.3.
- Text to videoPrompt to video, Pro quality tierOpen workflow
- Text to video (Fast)Prompt to video, faster and lower-cost tierOpen workflow
- Image to videoAnimate a still, Pro quality tierOpen workflow
- Image to video (Fast)Animate a still, faster and lower-cost tierOpen workflow
- Audio to videoDrive the picture from a supplied audio trackOpen workflow
- Extend a clipAdd 2 to 20 seconds onto an existing clipOpen workflow
- Retake a shotRegenerate a shot you already renderedOpen workflow
Cost is charged per second of output and scales sharply with resolution, and the Fast tier of each route is materially cheaper per second than Pro. Audio-to-video, extend and retake share one flat per-second rate. Credits come from one shared balance across image, video and speech, and the exact estimate is shown in the Studio before you generate.
Built for
- High-fidelity 4K content that has to hold up to close inspection
- Audio-driven video such as music videos and podcast visualisation
- Long-form work built from extend and retake passes
- LoRA fine-tuning for a custom character or house style
- Portrait 9:16 content for mobile feeds
FAQ
LTX-2.3, answered.
Limits, audio, licensing and price — the things worth knowing before you spend a credit.
Twenty seconds at 4K, on open weights.
Open the Studio, start with text to video, and see the credit estimate before you spend anything.