LTX, the company spun out of Lightricks, has released LTX-2.5, its latest open-weights video and "world" model, and launched it with day-one native integration in ComfyUI as part of a strategic partnership. The model is available on Hugging Face, inside ComfyUI, and via the LTX API for teams that prefer managed generation.
Usage is free for organizations with under $10 million in annual recurring revenue (ARR); larger enterprises negotiate licensing terms. LTX says its model family has reached 33 million downloads, which the company cites as evidence that it is the most-used open world model line on the market.
Key technical changes in LTX-2.5
According to LTX, LTX-2.5 rebuilds almost every stage of the generation pipeline rather than layering new features on an older core. The headline improvements include:
- A new diffusion video decoder that reduces artifacts in high-motion footage and better reconstructs fine detail such as text and faces while maintaining LTX's high compression ratio.
- Native multishot generation that renders a full sequence as a single output, preserving character, scene, and voice consistency across cuts rather than stitching separate shots together.
- A custom Gemma 4 language backbone and a dedicated prompt enhancer to better handle complex, multi-subject prompts.
- A pretrained checkpoint tuned for physical AI and robotics use cases, intended as a base for fine-tuning on domain-specific data that differs from cinematic video.
- A substantially improved distilled model that delivers near full-model quality at lower cost and faster inference; via an optimization effort with NVIDIA it can run locally on NVIDIA RTX GPUs with reduced memory requirements.
LTX says the model achieves roughly one-eighth the cost and one-seventh the render time of comparable models, and can run on hardware ranging from data-center GPUs down to Macs.
Performance and quality claims
The most prominent performance figure LTX cites is speed: it reports LTX-2.5 generates a 10-second, 720p image-to-video clip in 6.8 seconds — faster than real time. That measurement was taken on a self-hosted setup using two NVIDIA GB200 chips at steady state, a configuration many teams do not have. The same job run through LTX's managed API took 23.7 seconds, rendered at 1080p (the API has no 720p tier).
LTX also published end-to-end timings for competing APIs on the same task (as reported or commissioned by LTX): Google Gemini Omni Flash 52 s, xAI Grok 1.5 63 s, Google Veo 3.1 70 s (for an 8-second clip), MiniMax H3 180 s, ByteDance Seedance 2.5 317 s, and Kuaishou Kling 3.0 Pro 398 s. These numbers are vendor-reported and not independently verified.
On perceived output quality, LTX ran blind pairwise human preference tests. LTX-2.5 recorded a 67% win rate in those tests, narrowly ahead of Seedance 2.5 at 65%; other reported results were Gemini Omni Flash 55%, MiniMax H3 50%, Seedance 2.0 44%, Wan 2.6 42%, and FLUX 3 28%. LTX calls these preliminary results and recommends buyers validate models against their own workloads.
Deployment, licensing and business strategy
LTX emphasizes deployment flexibility: LTX-2.5 runs on any GPU with at least 16 GB of VRAM, can be deployed on-premises, at the edge, or via API, and does not require mandatory branding on generated outputs. The company contrasts this flexibility with closed API-only rivals and with open-licensed competitors whose weights are unavailable in certain regions or restrict fine-tuning.
Co-founder and CEO Zeev Farbman framed the release as part of a deliberate strategy favoring open weights over closed APIs. Farbman said many business use cases require access to model weights so teams can build custom flows, and that openness is not philanthropy but a business approach: free use for smaller organizations with licensing for larger customers.
Why ComfyUI matters
ComfyUI, which began as a January 2023 open-source side project, has become a widely used node-based prototyping environment in generative media. It is now backed by Comfy Org, which raised $17 million to continue development; Yoland Yan is a ComfyUI co-founder and CEO.
Farbman said the ComfyUI integration is important because many of LTX's paying customers discover and prototype workflows in Comfy first. Yan described ComfyUI as a connective layer in the open ecosystem that lets users run open-weight models locally, tap into closed models through partner nodes, and assemble workflows spanning creative pipelines, data processing, and robotics.
Both executives argued that ComfyUI acts as a customer acquisition channel for model providers: people prototype in ComfyUI, then enterprises purchase licenses when a workflow graduates to production.
Use cases beyond cinematic video
Farbman and Yan stressed that video and world models are used for more than making cinematic clips. Examples cited include computational photography and sensor noise reduction for hardware customers, VFX and environment effects in production studios, day-to-night transforms, and animation pipeline interpolation where animators provide keyframes and models interpolate between them.
Because LTX publishes open weights, experimentation can take place on a company's own hardware without per-generation billing and without data leaving the organization—an important consideration for enterprises handling sensitive footage, proprietary characters, or regulated data. The commercial trigger is scale: organizations with ARR above $10 million require a license.
Real-time, edge and developer scenarios
LTX positioned LTX-2.5 for real-time applications. In addition to ComfyUI, LTX named launch partners Asteria (an AI film studio) and Reactor (a developer platform that runs LTX-2.5 on low-latency inference infrastructure for interactive avatars, live worlds, and real-time robotics). These partners aim to let developers build production-grade, low-latency experiences without managing infrastructure themselves.
Yan pointed to Flipbook, a viral interactive experience on Reddit that generates a clickable world on the fly, as an example of what open weights plus low latency can enable; he said the entire interface is generated with an LTX model and live-streamed.
Farbman added that efficiency at the edge is a deliberate design goal for LTX, enabling deployment on consumer devices and edge hardware used for physical AI and robotics.
Filmmakers, studios and creators
Yan suggested that real-time world models are already altering virtual production workflows by narrowing the gap between shooting and post-production, giving producers and directors faster feedback. He cautioned, however, against literal reading of head-to-head comparisons because models specialize differently across animation, photorealism, gaming, 3D and robotics.
For indie and amateur creators, ComfyUI will continue to provide both approachable workflows and advanced lanes: Yan compared the experience to skiing, with easier slopes for newcomers and double-black-diamond capabilities for technical professionals.
Availability
LTX-2.5 is available now on Hugging Face as open weights, natively inside ComfyUI, and via the LTX API. LTX notes that its timing and preference measurements are preliminary and encourages prospective users to test the model on their own workloads.
