LTX-2.3 Model
LTX-2.3 Video Engine
Sharper detail. Cleaner audio. Stronger motion. Native portrait. Better video generation.
Sharper detail. Cleaner audio. Stronger motion. Native portrait. Better video generation.
Delivered as an IC-LoRA on LTX-2.3. Generate directly in HDR or convert existing SDR footage to EXR. More grading latitude, more range, ready for real finishing pipelines.
Generate vertical video up to 1080×1920 — trained on portrait-orientation data, not cropped from landscape.
Generate video where voice, music, and sound effects define structure, pacing, and motion.Built for production-grade workflows that require precise, harmonious control over audio-led scenes - from podcasts and avatars to voice-driven clips -not one-off demos or talking heads.
Extend creative range with long-form generation. Produce up to 20 seconds of high-fidelity video with complete control and consistent style.
Rebuilt latent space with an updated VAE trained on higher-quality data. Fine textures, hair, text, and edge detail are better preserved through the full generation pipeline.
4x larger text connector. Complex prompts — multiple subjects, spatial relationships, stylistic instructions — now resolve accurately. Try being more specific. The model handles it.
Less freezing, less Ken Burns, more real motion. Better visual consistency from the input frame. Fewer generations you throw away.
Filtered training data and a new vocoder. Fewer artifacts, fewer unexpected drops, tighter alignment across text-to-video and audio-conditioned workflows.
Generate vertical video up to 1080×1920, trained on portrait-orientation data, not cropped from landscape. Ready for production.
Delivered as an IC-LoRA on LTX-2.3. Generate directly in HDR or convert existing SDR footage to EXR. More grading latitude, more range, ready for real finishing pipelines.
Creative control that holds up under pressure. Structure, motion, camera behavior, and identity can be directed with intent rather than guessed by the model.
Depth-Aware Generation
OpenPose Driven Motion
Camera Control
Stylistic and Visual Consistency
Audio to Video
The model adapts to your worlds, characters, and creative DNA. Customization becomes part of the workflow, not a research project.
LoRA Training Support
Style LoRAs
Tools for upscaling, restoration, and detail recovery, powered by the model’s multi-scale rendering pipeline.
Detail Upscaling
Recreate and generate elements of already existing videos. Edit with surgical precision.
Retake
Extend Scene
Two flows, optimized for different production needs
Built for speed and tight feedback loops. Choose Fast Flow when rapid iteration matters more than maximum visual detail.
Technical characteristics:
High-fidelity generation for stable, detailed results. Choose Pro Flow when visual quality and consistency are more important than render speed.
Technical characteristics:
LTX built
Run LTX locally, integrate it into your stack, and build directly on the engine — full weights, full control, no lock-in.
Generate cinematic video directly from text prompts. Control motion, composition, and visual flow using natural language.
Animate still images into coherent video. Preserve visual identity while adding motion, transitions, and cinematic depth.
Edit and transform videos with precise control — refine scenes, enhance quality, and adjust motion while preserving continuity and character consistency.
Deployment
Run LTX-2.3 locally, integrate via API, or deploy at commercial scale — all powered by the same production-ready model.
Run LTX with full access to model weights and code. Deploy on your own infrastructure — fully local and on-prem ready, free for companies under $10M in annual revenue.
Access LTX 2.3 checkpoints via managed API endpoints. Integrate generation into your product or workflow without managing infrastructure or handling model deployment yourself.
For companies with over $10M in annual revenue embedding LTX into commercial products or production environments. Commercial terms tailored to your deployment model.
LTX built
The interface layer is a solved problem. What remains hard is the engine — the model that actually generates media. LTX is built to sit underneath whatever you want to create. We don't lock it behind a proprietary layer. We release it. We built on it too.
Run LTX locally from your terminal. Build pipelines, automate workflows, ship products.
Run LTX with full access to the model weights and code. Install, modify, extend, and deploy on your own infrastructure — fully local and on-prem ready.Free for companies with under $10M in annual revenue under the LTX Model License.
Access LTX 2.3 checkpoints via managed API endpoints. Integrate generation into your product or workflow without managing infrastructure.
For companies with over $10M in annual revenue embedding LTX into commercial products or production environments. The Licensing Program provides tailored commercial terms aligned with your deployment model and business needs.
LTX-2.3 is the latest release of the LTX-2 model family, a diffusion transformer (DiT) foundation model that generates high-fidelity video and synchronized audio from a single model. It supports text-to-video, image-to-video, and audio-to-video generation at up to 1080p, including native portrait (9:16) video. LTX-2.3 is available both as an open-source model and through the LTX API.
LTX-2.3 brings four major improvements over LTX-2.
A redesigned VAE produces sharper fine details, more realistic textures, and cleaner edges.
A new gated attention text connector means prompts are followed more closely — descriptions of timing, motion, and expression translate more faithfully into the output.
Native portrait video support lets you generate vertical (1080×1920) content without cropping from landscape.
And audio quality is significantly cleaner, with silence gaps and noise artifacts filtered from the training set.
Yes. LTX-2.3 model weights are freely available on HuggingFace under an open license. The release includes the base dev checkpoint, a quantized fp8 variant, and the distilled model for faster inference. Training code, ComfyUI custom nodes, and reference workflows are all available on the LTX-Video GitHub repository.
LTX-2.3 API usage is billed per second of generated video. Two model variants are available — ltx-2-3-fast for rapid iteration and ltx-2-3-pro for production-quality output — at both 720p and 1080p resolutions. Portrait and landscape generations are priced identically at the same resolution tier.
See the pricing page for current rates.
Yes, ComfyUI and Fal are fully supported. LTX-2.3 ships with updated ComfyUI custom nodes and reference workflows for text-to-video, image-to-video, and multi-stage generation with latent upscaling — available via the ComfyUI-LTXVideo repository.
Full documentation is available at docs.ltx.video, covering the API, open-source model setup, ComfyUI workflows, PyTorch integration, and usage guides for text-to-video, image-to-video, and portrait video. A migration guide for users upgrading from LTX-2 is also available there.
Yes. Multiple checkpoints are available for local use on HuggingFace, including the full dev model (bf16), a quantized fp8 variant for lower VRAM setups, the distilled model for faster inference, and spatial and temporal latent upscalers.
Yes — LTX-2.3 delivers sharper output, better prompt adherence, cleaner audio, and significantly improved image-to-video across the board. The one exception: if your workflow relies on custom LoRAs, those will need to be retrained for the 2.3 latent space before you migrate. See the Migration Guide for details.
