LTX, the open world model company spun out of Lightricks, has released LTX-2.5 — an open-weights video generation model that the company claims can produce a 10-second clip from a static image in just 6.8 seconds. The release marks a significant shift in the video AI market, where closed APIs from Google, OpenAI, and ByteDance have dominated.
The model arrives with native integration into ComfyUI, the popular node-based workflow tool that has become the de facto prototyping environment for open generative media. This day-one partnership reflects a broader strategy: betting that open weights — not closed APIs — will win the video and world model market.
“We’re trying to maintain the same efficiency and the inference speed that we’re known for, but constantly pushing the quality up,” said Zeev Farbman, LTX co-founder and CEO. The company reports 33 million downloads across the LTX model family, claiming it as the most-used open world model line on the market.
What’s New in LTX-2.5
The release rebuilds nearly every stage of the generation pipeline. Key improvements include a new diffusion video decoder that reduces visual artifacts in high-motion footage and reconstructs fine details like text and faces. Native multishot generation allows rendering a full sequence as a single output, maintaining character, scene, and voice consistency across cuts — a significant advancement over stitching individually generated shots together.
A custom Gemma 4 language backbone and dedicated prompt enhancer handle complex, multi-subject prompts more accurately. The release also includes a pretrained checkpoint tuned for physical AI and robotics, giving teams a base to fine-tune on domain data that differs substantially from cinematic video.
On cost, LTX-2.5 generates 720p video with audio at $0.09 per second through its API, putting a 10-second clip at $0.90. This undercuts premium models like Google’s Veo 3.1 ($4.00) and Black Forest Labs’ FLUX 3 Video ($1.70), though Google’s budget tier Veo 3.1 Lite at $0.50 remains cheaper per clip.
The Open Weights Play
The launch positions LTX against both closed API providers and open-licensed competitors whose weights are unavailable in the U.S. and Europe or restrict fine-tuning. Organizations under $10 million in annual recurring revenue can use LTX-2.5 free for self-hosting, with larger companies negotiating a license.
The model runs on any GPU with a minimum of 16GB of VRAM, deploying on-premises, at the edge, or via API. Notably, output carries no visible watermark, though the license requires disclosure that content is machine-generated and forbids removing embedded provenance features.
The speed benchmark was measured on two Nvidia GB200 superchips. Through LTX’s managed API at 1080p, the same job took 23.7 seconds — still notably faster than competitors, where Google’s Gemini Omni Flash measured 52 seconds on the same task.
Independent benchmarks remain limited. Gemini Omni Flash currently leads Artificial Analysis’ text-to-video arenas; LTX-2.5 has not yet been evaluated there. The company’s own human preference tests show a 67% win rate against Seedance 2.5’s 65%, though these figures are vendor-reported and labeled preliminary.