LTX-2.3: A Major Update to LTX-2 for AI Video Generation
Lightricks has released LTX-2.3, a significant upgrade to the LTX-2 video generation model. This new version introduces improvements across visual detail, motion stability, prompt understanding, and audio generation, making it one of the most capable open AI video models currently available.
The update builds on the powerful foundation of LTX-2 but refines several key areas of the generation pipeline, improving both video quality and overall reliability.
In this article, we’ll explore what’s new in LTX-2.3 and highlight a unified ComfyUI workflow that allows creators to generate videos using text prompts, images, or custom audio inputs.
For a complete walkthrough of the workflow and generation tests, watch the tutorial video included below.
Watch the Full LTX-2.3 Workflow Tutorial
The video tutorial demonstrates the complete workflow and generation process, including:
- Required models for LTX-2.3
- ComfyUI workflow structure
- Text-to-Video generation
- Image-to-Video animation
- Talking avatar generation with custom audio
- Multi-stage sampling and upscaling
Watch the full tutorial here:
What’s New in LTX-2.3
Sharper Fine Detail
One of the biggest improvements in LTX-2.3 is a rebuilt latent space and updated VAE trained on higher-quality data.
This change helps preserve fine visual details such as:
- Hair and skin textures
- Small objects and edges
- Text and graphical elements
The result is cleaner and more consistent outputs throughout the generation pipeline.
Improved Prompt Adherence
LTX-2.3 also improves how the model understands prompts by introducing a 4× larger text connector.
This allows the model to better interpret:
- Multiple characters or subjects
- Spatial relationships within scenes
- Complex stylistic instructions
Prompts that previously produced inconsistent results now resolve more accurately and predictably.
Better Image-to-Video Motion
Image-to-video generation has also been improved.
Compared to earlier versions, LTX-2.3 produces:
- More natural motion
- Better visual continuity from the input frame
- Fewer frozen frames or artificial camera movements
This makes it much more reliable for animating still images into dynamic scenes.
Cleaner Audio Generation
Another major upgrade is improved audio generation quality.
With filtered training data and a new vocoder, LTX-2.3 generates audio with:
- Fewer artifacts
- Better synchronization with video
- More stable speech generation
This improvement is particularly useful for talking avatars and audio-driven video workflows.
Native Portrait Video Generation
LTX-2.3 now supports native portrait video generation up to 1080 × 1920 resolution.
Unlike previous approaches that simply cropped landscape outputs, the model was trained on portrait-orientation data, allowing it to generate vertical video more naturally.
This makes it well suited for social media content and mobile-first video formats.
LTX-2.3 Model Versions
Similar to LTX-2, the new release includes two main model variants.
Dev Model (Maximum Quality)
The Dev model provides the highest generation quality and stability.
Typical settings include:
- CFG around 4
- 20 diffusion steps
This configuration prioritizes fidelity and accuracy.
Distilled Model (Maximum Speed)
The distilled model is optimized for faster generation.
Typical settings:
- CFG 1
- 8 diffusion steps
This version runs significantly faster while still producing strong visual results.
Distilled LoRA
Lightricks also provides a distilled LoRA for LTX-2.3.
Applying this LoRA to the Dev model allows it to run with:
- CFG 1
- 8 steps
This effectively combines the quality of the Dev model with the speed of the distilled setup.
New Upscaling Options
LTX-2.3 also introduces additional upscaling models designed for multi-stage video generation workflows.
These include:
- 1.5× Spatial Upscaler
- 2× Spatial Upscaler
- 2× Temporal Upscaler
The spatial upscalers allow progressive resolution increases across multiple stages, while the temporal upscaler increases frame rate directly in the latent space.
This approach allows creators to scale videos to higher resolution or higher FPS without regenerating the entire sequence.
Unified LTX-2.3 ComfyUI Workflow
To demonstrate how these capabilities can be used in practice, the tutorial video below features a unified ComfyUI workflow designed specifically for LTX-2.3.
This workflow supports multiple generation modes within a single pipeline.
Supported Generation Modes
- Text-to-Video (T2V)
Generate video directly from text prompts. - Image-to-Video (I2V)
Animate a starting image into a moving scene. - Custom Audio-to-Video with Lip Sync (A2V)
Use your own audio file to drive speech and lip-sync animation.
Flexible Model Support
The workflow also supports loading models in multiple formats:
- Safetensors models
- GGUF quantized models
This allows the workflow to run on a wide range of hardware configurations, from high-end GPUs to more limited systems.
Final Thoughts
LTX-2.3 represents a significant upgrade over LTX-2, bringing improvements across the entire AI video generation pipeline.
With better prompt adherence, sharper visual detail, improved motion, and cleaner audio, it offers a more reliable platform for building advanced video generation workflows.
When combined with a unified ComfyUI workflow supporting T2V, I2V, and audio-driven video generation, LTX-2.3 becomes an extremely flexible tool for creators exploring the next generation of AI video production.
If you want to see the workflow in action, be sure to check out the video tutorial above, where the full pipeline is explained step-by-step.