LTX-2.3 is an open-source generative video architecture based on the Diffusion Transformer (DiT). The model delivers commercial-grade generation quality (on par with Google Veo 3) but without the strict limitations of closed ecosystems. The project supports both fully autonomous local inference and hybrid cloud processing.
ltx-2.mp4
- Text-to-video generation
- Image-to-video generation
- Audio-to-video generation
- Video edit generation (Retake)
- Video Editor Interface
- Video Editing Projects
-
Commercial-Grade Quality: Spatiotemporal consistency, photorealism, and object physics are on par with closed proprietary models (Veo 3, Sora).
-
Flexible Deployment (Local / Cloud): * Local: Completely unlimited generation on your own hardware.
-
Cloud: Integrated cloud rendering for low-end PCs (queue wait times depend on server load).
-
Base Generation Parameters: Native output of 10 seconds at 24 FPS in 1080p resolution in a single pass.
-
4K Cloud Rendering: Integrated upscaling and native generation in 4K. A limit of 10 generations per day applies for cloud usage (to prevent server overload). Note: Video editing operations (V2V, Inpainting, Motion Brush) do not consume the 4K generation quota.
-
Spatiotemporal Attention &
$v$ -prediction: An optimized mathematical noise prediction model that minimizes visual artifacts in highly dynamic scenes. -
Fully Uncensored
| Feature | LTX-2.0 | LTX-2.3 (Current) |
|---|---|---|
| Architecture | DiT-Lite | Full-scale DiT with Spatiotemporal Attention |
| Base Resolution | 720p | 1080p (4K support) |
| Generation Length | 4-5 seconds | 10 seconds (seamless) |
| Prompt Adherence | Medium | High (understands multi-layered lighting and cinematography) |
| VRAM Consumption (Min) | 12 GB | 8 GB (thanks to new quantization algorithms) |
| Feature | LTX-2.3 | VEO 3 (Closed Source) |
|---|---|---|
| Diffusion Model | Open Source / Local / Cloud | Closed Cloud API |
| Limits (1080p) | Absolutely unlimited (Local) | Strict quotas (Credits) |
| Limits (4K) | Unlimited (Local) / 10 per day (Cloud) | Paid Tier |
| Physics Simulation Quality | High (Custom Motion Buckets) | High |
| Weights Access & Fine-tuning | Yes (LoRA, ControlNet) | No |
| Content Restrictions | None (Fully Uncensored) | Strict safety guidelines (Censored) |
LTX-2.3 utilizes dynamic CPU offloading and 8-bit/4-bit quantization, allowing the model to run smoothly on mid-range PCs. It won't run on a "potato," but basic modern systems are sufficient.
| Component | Minimum Requirements (Local) | Recommended Requirements (Local 4K) | Cloud Mode |
|---|---|---|---|
| OS | Windows 10/11, macOS 13+, Linux | Windows 11, Ubuntu 22.04 LTS | Any |
| CPU | 6-core (Intel i5 12th gen / AMD Ryzen 5) | 8-core+ (Intel i7 / AMD Ryzen 7) | Any |
| RAM | 16 GB DDR4 | 32 GB DDR5 | 8 GB |
| GPU | NVIDIA RTX 4060 (8 GB VRAM) | NVIDIA RTX 4080 / 4090 (16+ GB VRAM) | Not required |
| Storage | 25 GB (SSD) for model weights | 50 GB (NVMe SSD) for cache and video | 500 MB |
Note: macOS users (Apple Silicon M2/M3/M4) can use the Metal framework (MPS) for hardware acceleration. The expected generation time for a 10s 1080p video on an M3 Max is ~4-6 minutes.
The easiest way to get started is by using the pre-built binaries with an integrated UI. You do not need to install Python or set up virtual environments.
- Go to the Releases page.
- Download the installer for your OS:
- Windows:
LTX-2.3_x64.exe - macOS:
LTX-2.3_macOS.dmg
- Run the installer and follow the on-screen instructions. The installer will automatically download the necessary base model weights (approx. 18 GB).
When writing prompts, focus on detailed, chronological descriptions of actions and scenes. Include specific movements, appearances, camera angles, and environmental details - all in a single flowing paragraph. Start directly with the action, and keep descriptions literal and precise. Think like a cinematographer describing a shot list. Keep within 200 words. For best results, build your prompts using this structure:
- Start with main action in a single sentence
- Add specific details about movements and gestures
- Describe character/object appearances precisely
- Include background and environment details
- Specify camera angles and movements
- Describe lighting and colors
- Note any changes or sudden events
LTX-2 pipelines support automatic prompt enhancement via an enhance_prompt parameter.
To use our model with ComfyUI, please follow the instructions at ComfyUI-LTX-2.3.
Q: How fast is text-to-video generation? A: Speed depends on your hardware and the selected mode. When running locally on a flagship GPU (e.g., RTX 4090), generating a base clip (10 seconds, 1080p, 24 FPS) takes about 4–6 minutes. In cloud mode (Cloud Engine), the rendering itself takes 1-2 minutes, but there may be wait times depending on the server queue.
Q: How many GBs does the entire model weigh? A: The base optimized package (in FP8 format, which our installer downloads) weighs about 18–20 GB. This volume includes the diffusion weights (UNET), the autoencoder (VAE), and the text encoders (T5). The full uncompressed version (BF16) for server hardware takes up around 40 GB.
Q: Can it be integrated into ComfyUI?
A: Yes, the model works perfectly within the ComfyUI ecosystem thanks to its component-based structure. You will need custom nodes (e.g., ComfyUI-LTX-Wrapper), and the weights themselves are distributed across the /models/unet/, /models/clip/, and /models/vae/ directories. You can use our official .exe updater to automatically download and place all the necessary files ComfyUI-LTX-2.3.
Q: What graphics card is needed for a comfortable local launch? A: Thanks to quantization algorithms, the minimum threshold to run is 12 GB VRAM (e.g., RTX 3060 12GB), but generation will be slower due to partial data offloading to system RAM. The recommended capacity for comfortable and fast 1080p performance is 16–24 GB VRAM (RTX 4080, RTX 3090/4090).
Q: Are there any hidden usage limits? A: Local generation (Local Engine) is absolutely unlimited and free. Restrictions only apply to the cloud mode (Cloud Engine): base generation in 1080p is unlimited, but rendering in 4K resolution has a quota of 10 generations per day per account to prevent server overload.