LTX-2.5, a new open-weights video generation model released by Lightricks, generates 10-second clips in 6.8 seconds on top-tier hardware and undercuts most competitors on price, while claiming a 67% win rate in blind quality tests against rivals like Google's Gemini Omni Flash and ByteDance's Seedance 2.5. The model arrived on August 11, 2026, natively integrated into ComfyUI, a popular node-based workflow tool, marking a strategic shift toward open-source video generation over closed commercial APIs.

How Does LTX-2.5 Compare to Other Video Models?

The video generation market has become crowded with competing approaches to speed, quality, and cost. LTX-2.5 positions itself as the open-weights alternative to closed models, meaning developers can download the model weights for free if their organization has less than $10 million in annual revenue, or negotiate a license for larger companies. This contrasts sharply with Google's Veo 3.1, OpenAI's Sora, and ByteDance's Seedance 2.5, which operate as closed APIs accessible only through paid subscriptions.

On pricing, LTX-2.5 Fast tier costs $0.09 per second, or $0.90 for a 10-second clip at 720p resolution with audio. This undercuts Google's Veo 3.1 Fast ($1.00 per 10-second clip) and Gemini Omni Flash ($1.00), but sits above Google's budget option, Veo 3.1 Lite, which costs $0.50 per clip. For comparison, FLUX 3 Video costs $1.70 per 10-second clip, Alibaba's HappyHorse 1.0 runs roughly $1.82, and Kuaishou's Kling 3.0 Pro does not publish per-second pricing.

The speed advantage is more dramatic when measured on specialized hardware. LTX-2.5 generates a 10-second, 720p clip in 6.8 seconds on two NVIDIA GB200 superchips, a configuration far beyond what most teams can access. Through LTX's own managed API, the same task takes 23.7 seconds at 1080p resolution. For context, Google's Gemini Omni Flash takes 52 seconds, Google's Veo 3.1 takes 70 seconds, and ByteDance's Seedance 2.5 takes 317 seconds on the same benchmark.

What Technical Improvements Does LTX-2.5 Introduce?

Rather than adding features to an older foundation, LTX rebuilt nearly every stage of its generation pipeline. The company introduced several key upgrades designed to improve both quality and usability for different applications:

• Diffusion Video Decoder: A new decoding stage that reduces visual artifacts in high-motion footage and reconstructs fine details like text and faces while maintaining LTX's compression efficiency.
• Native Multishot Generation: The ability to render a full sequence as a single output, keeping character, scene, and voice consistent across cuts instead of stitching individually generated shots together.
• Improved Language Backbone: A custom Gemma 4 language model and dedicated prompt enhancer that handles complex, multi-subject prompts more accurately.
• Robotics-Tuned Checkpoint: A pretrained version optimized for physical AI and robotics applications, giving teams a foundation to fine-tune on domain-specific data.
• Distilled Model Option: A smaller, faster version that delivers near-full-model quality at lower cost and can run locally on NVIDIA RTX GPUs with reduced memory requirements.

The distilled model represents a significant practical advantage for teams without access to data center resources. It enables local deployment on consumer-grade hardware, reducing latency and eliminating API dependency for time-sensitive applications.

How Does Quality Compare in Independent Testing?

LTX commissioned blind, side-by-side human preference tests in which evaluators voted on videos generated from the same prompts without knowing which model produced them. LTX-2.5 recorded a 67% win rate, narrowly ahead of ByteDance's Seedance 2.5 at 65%, with Google's Gemini Omni Flash at 55%, Alibaba's MiniMax H3 at 50%, and Black Forest Labs' FLUX 3 at 28%.

However, these results come with an important caveat: they were commissioned by LTX and have not been independently verified. The company itself labels the preference results as preliminary and expects them to evolve as evaluation expands. Independent benchmarks tell a different story. As of August 2026, Google's Gemini Omni Flash leads both of Artificial Analysis' text-to-video arena leaderboards, placing it ahead of LTX-2.5 in third-party evaluations.

"We are introducing many cool things in this release: multi-shot support, a diffusion decoder for better quality, new conditioning modes, better support for autoregressive models that are critical for real-time use cases and robotics," said Zeev Farbman, co-founder and CEO of LTX.

Zeev Farbman, Co-founder and CEO at LTX

Why Does the Open-Weights Strategy Matter?

LTX and ComfyUI are betting that open-weights models will win the video generation market over closed APIs. The LTX family has already passed 33 million downloads, making it the most-used open-weights world model line available. This strategy offers developers several advantages: the ability to fine-tune models on proprietary data, local deployment without API calls, and no vendor lock-in.

The day-one integration with ComfyUI underscores this commitment. ComfyUI has become the de facto prototyping environment for open generative media, and native support means developers can access LTX-2.5 directly within their existing workflows without switching tools or learning new interfaces. For organizations under $10 million in annual revenue, the model is free to use, lowering the barrier to entry compared to closed APIs that charge per-second or per-minute rates.

The trade-off is clear: open-weights models require more technical expertise to deploy and optimize, while closed APIs offer simplicity and managed infrastructure. For teams with engineering resources and domain-specific use cases, the open approach provides flexibility. For teams prioritizing ease of use and support, closed APIs remain the simpler choice.

What Does This Mean for the Video Generation Market?

LTX-2.5's release signals a maturing market where speed, cost, and quality are converging across multiple competitors. No single model dominates all three dimensions. Google's Veo 3.1 Lite remains the cheapest option at $0.50 per clip. Gemini Omni Flash leads independent quality benchmarks. LTX-2.5 offers the only open-weights option with competitive speed and quality, plus the ability to extend clips up to 20 seconds at 24 or 25 frames per second.

The emergence of open-weights alternatives like LTX-2.5 may pressure closed-API providers to improve pricing or transparency. Conversely, closed providers can invest in quality improvements and user experience that open models may struggle to match. The market is likely to stratify: cost-conscious teams will gravitate toward budget tiers, quality-focused teams will use Gemini Omni Flash or LTX-2.5 Pro, and teams needing extended clips or specialized features will choose models like Veo 3.1 or Seedance 2.5.