Autoregressive Streaming Architecture Powers Real-Time, Long-Form AI Video with 8X Greater Length Than Industry Standard
NEW YORK, July 16, 2025 /PRNewswire/ — Lightricks, the company behind LTX Video (LTXV) and LTX Studio and a global leader in generative AI video innovation, announced today a major evolution in its AI video technology: an update to LTXV that enables the generation of clips longer than 60 seconds. This milestone positions Lightricks as the first company to enable live-streamed, long-form AI video creation at scale. This breakthrough represents an eightfold leap beyond the current industry norm of eight (8) seconds, and is the first real-time, streaming-capable, long-form AI video model in production today. Unlike traditional models limited to short outputs, LTXV’s autoregressive streaming architecture continuously generates video in real time, enabling developers and creators to build longer, coherent stories – not just isolated clips.
60 second AI video of frog on branch in the rain
As with prior LTXV updates, this new capability will maintain open weights and be available to developers, academia, and gen AI video enthusiasts. It will also power commercial products like LTX Studio, Lightricks’ flagship creative development platform used by creators and media teams to design, iterate, and produce video projects using AI.
The new LTXV release introduces an improved autoregressive video engine that allows video clips to stream live to viewers as they are rendered. The system returns the first second of content almost instantly and continues generating forward in real time, enabling interactive, continuous narratives with full control over scene development. This paves the way for a new category of generative storytelling applications, from player-generated cutscenes and adaptive educational content to real-time AR visuals synced with live performers.
“Crossing the 60-second mark unlocks a new era for generative media,” said Zeev Farbman, co-founder and CEO of Lightricks. “LTXV is unique in its ability to create longer scenes while maintaining full control of the extended sequences, which enables coherent storytelling with visual and semantic consistency, transforming AI video from a demo or just a random clip, into a true medium with creative intent.”
LTXV’s autoregressive architecture supports both Lightricks’ 13B and mobile-friendly 2B parameter models. Creators and developers can apply pose, depth, or canny control LoRAs not only at the beginning of a prompt but continuously throughout the 30-second-plus scene. Compatible with Lightricks’ IC-LoRA infrastructure, the system also enables near real-time motion capture feeds, expanding its usability across interactive platforms.
“We’ve reached the point where AI video isn’t just prompted, but truly directed,” added Yaron Inger, co-founder and CTO. “This leap turns AI video into a longform storytelling platform, and not just a visual trick.”
Technical Highlights:
Real-time autoregressive sequence conditioning across model variants (video generated in chunks of frames, with each chunk conditioning the generation of the next). This allows motion and narrative to be built with smooth continuity, much like a writer creates a story sentence by sentence, each informed by the last.
Cost and Efficiency: LTXV runs efficiently on a single H100 or even a consumer-grade GPU, while delivering artifact-free 30-second clips. In contrast, public benchmarks for competing solutions show significantly higher compute requirements, with multiple H100s needed to generate just a 5-second 1080p clip (for example, 41 seconds on an L20 and up to eight H100s for higher-resolution real-time output).
Streaming-first architecture: approximately one (1) second returned instantly, remainder streams live
Supports continuous control inputs for dynamic scene generation
Fully compatible with IC-LoRA motion and style LoRA integrations
Speed – first second returned in ~1 sec, full 60 seconds in real time (builds on the streaming architecture already described)
Key Use Cases Include:
Advertising and Social Media: 15 to 60-second vertical ad spots generated on demand
Gaming: Live-rendered cutscenes generated from gameplay data
Live Events: Stage-synced AR characters reacting in real time
Education: Adaptive explainer videos that evolve with learner input
LTXV is available as an open weights model on both Hugging Face (LTX-Video) and GitHub (LTX-Video) and already fully integrated into Lightricks’ flagship storytelling platform, LTX Studio. With a growing library of models designed for diverse creative needs and a commitment to open development, Lightricks is shaping the future of generative AI video, bridging research-driven breakthroughs with real-world application. For more information about Lightricks, its products, technology, and open source initiatives, visit www.lightricks.com.
SOURCE Lightricks
