The nascent generative AI video industry is turning into a real battleground, and Lightricks has upped the ante with the first long-form video generation model, which lets creators generate up to 60 seconds of content from a single prompt.
It’s a major breakthrough for generative AI video and a real statement of intent from Lightricks, as the video length permitted overshadows that of its competitors; the industry average is around eight seconds.
Extended output isn’t the only innovation in the latest release. In addition, LTX Video, also known as LTXV, is the first such model to support real-time streaming of AI-generated outputs. Creators can instantly stream their AI video as it’s being rendered – before it’s finished generating. Users will be able edit their content interactively, adding fresh prompts to dynamically update the content along the way.
In a press release announcing the model update, Lightricks termed the latest version of LTXV the world’s “first real-time, streaming-capable, long-form AI video model in production today,” and it’s an update that reinforces the belief that such tools are no longer just for enthusiasts, but are becoming a viable alternative to traditional video for professional content creators.
Lightricks CEO Zeev Farbman called the release a “new era” for generative media, saying that these tools are no longer just for creating random video clips, but are a “true medium for creative intent.”
He pointed to LTXV’s capabilities, which, for the first time, allow creators to maintain “full control of extended sequences.” He says it’s the first generative AI video model to properly enable “coherent storytelling with visual and semantic consistency.”
Livestreamed, real-time video generation
The secret sauce in Lightricks’ improved LTXV is an “autoregressive video engine,” which lets creators tinker with their videos while they’re being generated and streamed over the web. Users can keep prompting the model – for instance, by telling it to adjust the camera angle, or giving new instructions to the characters and objects on screen – essentially altering the story in real time. In this way, LTXV becomes an interactive, real-time storytelling tool unlike any other.
“We’ve reached the point where AI video isn’t just prompted, but truly directed,” said Lightricks CTO Yaron Inger.
These capabilities enable some intriguing new applications for generative AI video that would previously have been impossible, Lightricks states. For instance, the new model could be used to generate live cutscenes in video games based on a gamer’s controller inputs. Alternatively, it could be used to generate new augmented reality experiences in real-time, like AI-generated dancers that mimic the moves of Shakira during a livestreamed concert. Educational content is another possibility, with Lightricks floating the idea of tutorial videos that respond in real-time based on students’ inputs.
AI video’s fast-moving evolution
The release differentiates Lightricks when rivals are also releasing innovative features. One competitor is the Los Angeles-based AI video-generation startup Moonvalley, which last week launched the Marey model, notable for its ability to mimic motion and apply it to different objects and environments.
For instance, a creator could upload a video of a herd of buffaloes racing through grassland and superimpose an AI-generated sports car onto the same environment. The grass and dirt would respond in the same way, but it would appear as if the car is affecting the scene.
Moonvalley states Marey is the first AI video model trained entirely on open-source data, meaning there are no licensing issues with the content generated. Lightricks, notably, trained LTXV on content licensed from Getty Images and Shutterstock.
As impressive as Moonvalley’s and Lightricks’s models are, neither of them can do what Google’s latest AI video model Veo 3 does. Launched in May, Veo 3 became the first – and so far only – model of its kind to generate audio alongside video. Its comprehensive audio capabilities include everything from character voices to animal noises, sound effects, and musical soundtracks.
Google has also launched a new tool, Flow, aimed at filmmakers. It allows them to generate “cinematic videos” simply by describing a location, the desired camera angle, and style preferences. Flow can be accessed through Google’s AI assistant Gemini, Google Workspace, and the Vertex AI platform.
Alibaba has launched Wan 2.1-VACE, an open-source video model that offers editing tools and supports multimodal inputs including text, images, and video.
Midjourney, regarded as one of the best AI image generators, has also entered the video generation space with V1, released in June. Midjourney’s V1 is an image-to-video model that can generate five-second clips; limited, but it underscores the competition present in the AI video space.
Plenty to like
Lightricks will pay close attention to its competitors, but the company’s evolving suite of apps and models is intriguing. Creators and content shops willing to pay a subscription fee can access the latest version of LTXV through LTX Studio, a professional video editing platform that provides various tools for users to ideate, flesh out, and refine their AI-generated videos.
Unlike many competing models, LTXV is open-source and can be downloaded from Hugging Face and GitHub, complete with open weights. Users are free to experiment with, and if they’re so inclined, adapt the model by fine-tuning it for their own workflows.
The LTXV model is designed to be run on low-power hardware, with Lightricks claiming that the 13 billion parameter version only needs a single H100 GPU to produce high-quality video outputs.
