Lightricks, the company behind the popular Facetune app, has released LTX-2.5, an open-weights 22-billion-parameter video generation model on August 11, 2026, in partnership with NVIDIA's Local AI Community initiative. The model introduces a breakthrough feature: native multishot generation — the ability to produce multiple connected scenes in a single pass while maintaining character identity, lighting, voice, and style across cuts.
Source: Unsplash — representing AI video production concept
What Is LTX-2.5?
LTX-2.5 is an open-weights Diffusion Transformer (DiT) model from Lightricks, designed for local execution and fine-tuning. It generates high-fidelity synchronized video and audio from text, images, and video inputs. A research paper titled "LTX-2: Efficient Joint Audio-Visual Foundation Model" was published on arXiv in January 2026.
The model is available under the LTX-2 Community License, which is free for entities with annual revenue under $10 million, with a paid license required for larger organizations.
Revolutionary Features
Native Multishot Generation
This is the headline feature: instead of generating a single continuous shot, LTX-2.5 can produce multiple connected scenes in a single generation pass. The model maintains character identity, lighting, audio, and visual style across consecutive cuts. This means you can write a simple three-scene script and get a coherent video that tells a complete story.
Custom Gemma 4 12B Text Encoder
The model uses a custom Gemma 4 12B text encoder with projection layers, specifically designed to hold complex prompts (multiple characters, camera movements, lighting setups) without dropping details.
Optimized for NVIDIA RTX
The model is available in multiple formats:
- BF16: Full quality for training and fine-tuning
- INT8: For constrained hardware (ComfyUI only)
- NVFP4: For Blackwell GPUs
On an RTX 6000 PRO, the model achieves 20% faster performance and 40% memory savings.
New Diffusion Video Decoder
LTX-2.5 replaces the traditional VAE decoder with a new diffusion video decoder that produces sharper faces, better textures, clearer on-screen text, and smoother motion with fewer artifacts. This is an important technical improvement because the VAE decoder was a weak point in previous video generation models.
Dynamic Compute Allocation
The "diffusion fidelity rendering" feature dynamically allocates compute power based on scene complexity and available budget. This means simple scenes generate quickly, while complex scenes (multiple characters, complex lighting) receive additional computation to ensure quality.
Duration Predictor
An optional feature that automatically predicts clip length from the text prompt, eliminating the need to manually specify frame counts. This is useful for users who are not familiar with the technical details of video generation.
Prompt Enhancer
The model includes a prompt enhancer that expands short prompts into richer cinematic instructions, helping users who may not be skilled at writing complex prompts.
What This Means for You
For Content Creators
Produce Complete Stories: The multishot feature means content creators can produce a short video that tells a complete story — not just a single animated shot. This is perfectly suited for Instagram Reels and YouTube Shorts that are growing in popularity.
Audio and Video Together: Like MiniMax-H3, LTX-2.5 generates audio and video together, but with the added benefit of multishot capability.
Full Customization: Being open-weights and fine-tunable, you can train the model on specific visual styles that suit your audience.
For Developers and Businesses
Local Execution: The model runs entirely locally, ensuring data privacy and full control over outputs.
ComfyUI Integration: The model integrates seamlessly with ComfyUI, allowing it to be incorporated into existing production workflows.
Programmable API: Beyond ComfyUI, the model provides a Python library (ltx-pipelines) for direct execution and a diffusers pack for programmatic integration.
Image: Unsplash — representing multishot generation concept
Quick Comparison: LTX-2.5 vs Competitors
| Feature | LTX-2.5 | MiniMax-H3 | Sora | Runway Gen-3 |
|---|---|---|---|---|
| Multishot | Yes (native) | No | No | Limited |
| Sync audio | Yes | Yes | Yes | No |
| Open weights | Yes | Yes | No | No |
| Runs locally | Yes | Yes | No | No |
| Size | 22B | 33B | Closed | Closed |
| Fine-tuning | Yes | Limited | No | No |
| License | LTX-2 Community | Open | Closed | Closed |
Both LTX-2.5 and MiniMax-H3 are open and run locally, but LTX-2.5 stands out with native multishot generation and full fine-tuning capability, while MiniMax-H3 has a larger parameter count and wider adoption.
How to Use It
Technical Requirements
- GPU: NVIDIA RTX with at least 16GB VRAM (24GB recommended)
- RAM: 32GB
- Storage: Approximately 50GB for the full BF16 model, less for INT8 and NVFP4
- Python: 3.12 or newer, CUDA 12.7+, PyTorch 2.7+
Getting Started
- Log in to HuggingFace and accept the model terms (gated model)
- Download model files via
hf download - Use ltx-pipelines for Python or ComfyUI for a graphical interface
- Start with simple prompts and let the prompt enhancer improve them
Company Context and Technical Landscape
Lightricks is no stranger to the digital content world. The Israel-based company developed the popular Facetune app (over 200 million downloads) and a suite of photo and video editing applications. The launch of LTX-2.5 represents a strategic shift from content editing applications to generation infrastructure.
The research paper published in January 2026 (arXiv:2601.03233) titled "LTX-2: Efficient Joint Audio-Visual Foundation Model" demonstrates serious research investment in this space. The model is built on a Diffusion Transformer architecture and leverages advanced techniques including:
- DiffVAE and Conv VAE: Two different systems for encoding and decoding video — DiffVAE for high quality and Conv VAE for speed
- Audio VAE with vocoder: A specialized system for audio processing and generation
- Spatial and temporal latent upscalers: Tools for upscaling video resolution both spatially and temporally
This integrated technical architecture is what enables LTX-2.5 to deliver the multishot feature — the model doesn't generate separate shots but understands visual narrative as a whole.
Comparison with Previous Versions
LTX-2.5 is a major evolution from the original LTX-2. Key improvements include:
- Multishot support not available in the previous version
- New prompt enhancer built on Gemma 4
- New diffusion video decoder
- Improved distilled model retaining higher quality in a smaller footprint
Limitations
- Gated Model: Requires HuggingFace login and approval to download, adding an extra step.
- Enterprise License: Companies with revenue above $10 million need a paid license.
- Hardware: Running the full model (BF16) requires powerful hardware. Lighter formats (INT8, NVFP4) reduce requirements but at a quality cost.
- Quality vs Closed Models: While LTX-2.5 offers excellent quality for an open model, closed models like Sora may outperform it in some scenarios.
Frequently Asked Questions
Is LTX-2.5 free?
The model is free for entities with annual revenue under $10 million under the LTX-2 Community License. Larger companies need a paid license.
What is the difference between LTX-2.5 and MiniMax-H3?
LTX-2.5 offers native multishot generation and full fine-tuning capability, while MiniMax-H3 focuses on video and audio quality with a larger size (33B vs 22B).
Does it support Arabic text prompts?
Yes, Arabic prompts can be used, but optimal performance is with English prompts. It is recommended to translate and then pass the English text to the model.
Can I use it for commercial content?
Yes, within the terms of the LTX-2 Community License for small and medium businesses. For large companies, a paid license is required.
What is the best GPU for running it?
For optimal BF16 performance: RTX 4090 or RTX 5090. For practical use with NVFP4: RTX 4070 Ti is sufficient.
Conclusion
LTX-2.5 from Lightricks is a powerful addition to the open AI video generation ecosystem. The multishot generation feature is what truly sets it apart — the ability to produce multi-cut video with full continuity in a single pass is a game-changer for content creators. With full ComfyUI support and local execution capability, the model offers a compelling alternative to closed paid services.
For more AI tools for creators, visit our AI Tools Guide on Truescho.