LTX-2.3: Lightricks Releases Open Source 4K Video AI Model

Until this year, generating broadcast-quality video with AI meant either paying Sora's per-second rates or accepting that open source tools would lag behind by 12 to 18 months. That gap closed on March 5, 2026, when Lightricks released LTX-2.3 — a 22 billion parameter, Apache 2.0-licensed model capable of producing native 4K video at 50 frames per second with synchronized ambient audio, all in a single inference pass. The model weights are free to download from Hugging Face and the full codebase is on GitHub.

This is the latest step in a fast-moving story. Lightricks launched its original LTX Video model in November 2024, scaled it to 13 billion parameters in May 2025, and then fully open-sourced the LTX-2 family in January 2026 before releasing this updated LTX-2.3 checkpoint two months later. For developers, independent filmmakers, and creative studios who need video generation without usage-based billing, this matters — a lot.

What exactly is LTX-2.3 and how does it work?

LTX-2.3 is a Diffusion Transformer (DiT) model — the same architectural family as Sora and Stable Diffusion 3 — but purpose-built for joint audio-video generation. Unlike earlier open source video tools that generated silent clips and bolted on audio as an afterthought, LTX-2.3 treats audio and video as a single unified output. Ambient sound, environmental effects, and motion are generated together in one model pass, which is why the audio stays temporally consistent with what's happening on screen.

The model accepts text prompts, image prompts (image-to-video), or combinations of both. It ships in four checkpoint variants on Hugging Face to support different use cases:

• ltx-2.3-22b-dev — full bf16 weights for fine-tuning and training; around 42 GB
• ltx-2.3-22b-distilled — 8-step fast inference, CFG=1; ideal for iteration
• ltx-2.3-22b-distilled-1.1 — updated distilled variant with improved audio rendering
• Spatial and temporal upscalers — separate 2x upscaler checkpoints for multi-stage pipelines targeting higher resolution or frame rate

For most production use cases, the fp8 quantized build delivers roughly 90% of the full-quality output at about 18 GB — achievable on an RTX 4080 or RTX 4090 without resorting to CPU offloading. According to Lightricks, all training data was licensed from Getty Images and Shutterstock, removing the legal ambiguity that hangs over most open source video models. This is a significant commercial differentiator.

What hardware do you actually need to run LTX-2.3?

Lightricks has been clear about the minimum and recommended specs. CUDA (NVIDIA) is the only officially supported GPU platform at launch, though community forks for ROCm (AMD) and MLX (Apple Silicon) are in progress. Here is a practical breakdown:

One non-obvious constraint: resolution settings must be divisible by 32, and frame counts must be divisible by 8 plus 1. If your prompt targets non-standard dimensions, you will need to pad with -1 then crop to the desired output. This is a minor workflow inconvenience but worth knowing before you build a pipeline around it. The codebase requires Python 3.12+, CUDA 12.7+, and PyTorch 2.7. For Mac users, the Lightricks API is currently the most reliable path — local MLX support is experimental and significantly slower.

How does LTX-2.3 compare to Sora, Runway, and other alternatives?

The March 2026 landscape for AI video generation is competitive, but LTX-2.3 occupies a unique position: it is the only model offering native 4K resolution at 50 FPS that you can self-host for free. Here is how the major players stack up for practical production use:

In benchmark rankings from Artificial Analysis, LTX-2 ranked in the top three for image-to-video creation, behind Kling 3.5 and Veo 3.1 by Google. The text-to-video benchmark placed it seventh. Those rankings reflect a single point in time — given the pace of open source development, LTX-2.3's positioning will likely shift as community fine-tunes and LoRA adapters mature over 2026.

What can you actually build with LTX-2.3 right now?

The use cases fall into three broad categories: creative production, commercial applications, and research. For creative production, the ComfyUI integration ships out of the box with reference workflows for text-to-video, image-to-video, and multi-stage generation with latent upscaling. If you already have a ComfyUI setup for image generation, adding LTX-2.3 requires downloading the Lightricks custom node package and the relevant checkpoint — no new infrastructure needed.

For commercial applications, the Apache 2.0 codebase and permissive weights license mean you can embed LTX-2.3 into a product, white-label it, or build a SaaS on top of it without licensing fees. The Getty/Shutterstock training data provenance removes the copyright-infringement risk that has made enterprises cautious about deploying open source image and video models. Several startups have already built production SaaS templates on top of the API, and Lightricks has made its own hosted API available at approximately $0.04 per second for Fast mode — around five times cheaper than comparable closed model pricing.

For research, the ltx-trainer included in the repository enables fine-tuning on custom datasets. This is where the long-term value lies: domain-specific fine-tunes trained on medical imaging, satellite footage, product photography, or architectural visualization could produce models that out-perform the general base on narrow tasks. The open source AI landscape has consistently shown that fine-tuned specialized models close the quality gap faster than most expect.

A few constraints worth flagging: the synchronized audio excels at ambient sounds and environmental effects but does not yet compete with dedicated music generation or voice synthesis tools. Think of it as automatic foley generation rather than full audio production. The image-to-video mode occasionally produces frozen frames or slow pans in edge cases involving complex physics — water simulations and crowd scenes are the most common failure modes. LTX-2.3 has improved this over LTX-2, but it has not been fully eliminated. Official Diffusers library support is still listed as coming soon; if your pipeline is tightly integrated with Diffusers, factor in additional integration time.

How do you get started with LTX-2.3 for free?

There are three routes depending on your hardware situation and workflow preference. The fastest way to try the model without any local setup is the LTX API Playground, which provides free trial credits and supports both text-to-video and image-to-video in the browser. This is ideal for evaluating the model before committing to local infrastructure.

For local deployment, the official path is:

• Clone the repository: git clone https://github.com/Lightricks/LTX-2.git
• Install dependencies with uv sync (requires Python 3.12+ and CUDA 12.7+)
• Download the model weights from Hugging Face — choose the distilled checkpoint for fast iteration or the dev checkpoint for fine-tuning
• Run inference via the provided CLI or Python scripts; ComfyUI workflows are included for node-based generation

LTX Desktop is Lightricks' downloadable GUI application that wraps the model in a consumer-friendly video editor. It launched alongside LTX-2.3 in March 2026 and is the recommended path for creators who want local generation without touching a terminal. The desktop app handles model download, quantization selection, and basic prompt management. For team deployments and self-hosted production pipelines, the Docker-based on-premises setup documented in the repository is the more scalable path. You can read more about evaluating free vs paid AI tools to understand where open source self-hosting makes sense for your workflow.