LTX, the open world-model company spun out of Lightricks, released LTX-2.5 on August 11, 2026, an open-weights video and world model that generates a 10-second, 720p image-to-video clip in 6.8 seconds on two NVIDIA GB200 chips. It arrives with day-one native integration in ComfyUI, the node-based workflow tool that began life as the Stable Diffusion community's favorite front end.
The model is live now on Hugging Face, inside ComfyUI, and through the LTX API for teams that want managed generation. It is free for organizations under $10 million in annual recurring revenue; larger companies negotiate a license. LTX says its model family has passed 33 million downloads, making it the most-used open world-model line on the market.
What's new in LTX-2.5
Rather than bolting features onto the old core, LTX-2.5 rebuilds nearly every stage of the generation pipeline, according to the company. The headline changes:
- A new diffusion video decoder that cuts visual artifacts in high-motion footage and reconstructs fine detail such as text and faces while keeping LTX's high compression ratio.
- Native multishot generation that renders a full sequence as a single output, holding character, scene, and voice consistent across cuts instead of stitching shots together.
- A custom Gemma 4 language backbone plus a dedicated prompt enhancer for complex, multi-subject prompts.
- A pretrained checkpoint tuned for physical AI and robotics, giving teams a base to fine-tune on domain data that looks nothing like cinematic video.
- A substantially improved distilled model that, through an optimization effort with NVIDIA, runs locally on RTX GPUs with reduced memory requirements.
LTX claims roughly one-eighth the cost and one-seventh the render time of comparable models, on hardware ranging from data center GPUs down to a Mac, with a minimum of 16GB of VRAM.
How the numbers stack up
Speed is the headline number. The 6.8-second figure was measured self-hosted on two of NVIDIA's top-end GB200 chips at steady state, a configuration far beyond what most teams have in a rack. The same job through LTX's own managed API took 23.7 seconds, albeit rendered at the higher 1080p resolution, since the API has no 720p tier.
By LTX's end-to-end measurements of competing APIs on the same task, Google's Gemini Omni Flash came in at 52 seconds, xAI's Grok 1.5 at 63 seconds, Google's Veo 3.1 at 70 seconds for an 8-second clip, MiniMax H3 at 180 seconds, ByteDance's Seedance 2.5 at 317 seconds, and Kuaishou's Kling 3.0 Pro at 398 seconds.
On quality, LTX shared results from blind, side-by-side human preference tests, in which evaluators voted on videos generated from the same prompt without knowing which model produced which. LTX-2.5 recorded a 67% win rate, narrowly ahead of Seedance 2.5 at 65%, with Gemini Omni Flash at 55%, MiniMax H3 at 50%, Seedance 2.0 at 44%, Wan 2.6 at 42%, and FLUX 3 at 28%.
All of these figures are vendor-reported, not independently verified, and LTX labels the preference results preliminary. They are directional claims a buyer should test against their own workloads.
Why the ComfyUI partnership matters
LTX CEO Zeev Farbman was unusually candid about the motivation: a whole lot of LTX's customers start their journey inside Comfy, so day-zero support for the integration is effectively the company's customer acquisition channel. ComfyUI began in January 2023 as an open-source side project by a pseudonymous developer known as comfyanonymous, who built a node-based graphical interface for Stable Diffusion that let users chain models and processing steps into repeatable visual workflows. It has since become the standard environment where new image and video models are tested and pushed into production, backed by Comfy Org, which raised $17 million.
Farbman was equally blunt about the open-weights strategy. The company started building its own models out of necessity when the big players began closing theirs, and he argues video and world models have a fundamentally wider surface area of use cases than language models, which is why access to the weights matters. He stressed that openness is not charity: individuals and companies below the revenue threshold use the model free, and successful teams eventually sign a licensing agreement.
For enterprises, the pitch is that experimentation can happen on their own hardware, with no per-generation billing and no data or IP leaving their systems, a meaningful distinction for studios with sensitive footage or regulated data. The commercial trigger only arrives with scale: organizations above $10 million in annual recurring revenue need a license.
Beyond ComfyUI, LTX named two other launch partners: Asteria, an AI film studio producing original film and video on LTX, and Reactor, a developer platform that runs LTX-2.5 on low-latency inference infrastructure to power interactive avatars, live worlds, and real-time robotics workloads.
The release lands as the open-weights camp pushes harder against closed API rivals. With LTX-2.5, LTX is betting that video and world models have too many use cases for closed APIs to win, and that builders will pay once they scale past the free tier.
Comments