New Delhi: IndiaAI Mission-backed Varya a distilled video‑generation model was launched by Avataar, an AI-native transformation company on Friday “to make frontier video AI affordable, accessible and relevant for India’s next generation of users”.
Varya uses a distillation technique that reduces video generation from 50 steps to 4 steps and generates video at Rs 0.48 per second, making it up to 10 times more cost-efficient than several leading global video models, the statement from Ministry of Electronics & IT said.
Avataar was among the companies selected by the IndiaAI Mission to build indigenous foundation AI capabilities. Access to subsidized national AI compute infrastructure enabled the research that led to Varya, the ministry said.
“The product experience is designed around a simple promise: Idea → Video → Story. Users can type an idea, upload an image, generate a video, and continue the story through additional clips,” the statement said.
One prompt can become a lesson, an ad, a guide, a film or a memory, the ministry noted.
The launch was conducted here in the presence of S Krishnan, Secretary, Ministry of Electronics and Information Technology (MeitY), alongside Avataar’s leadership team.
Designed for Indian contexts, Varya is built to understand and generate culturally rich visual outputs across India’s regions, festivals, communities, food, clothing, public spaces and everyday life, the statement said.
As video becomes the primary medium for learning, commerce, communication and storytelling, the ability to generate high-quality video content efficiently is becoming increasingly important.
“Through strategic support for foundational models, we are enabling innovation at scale and creating the building blocks for the next generation of AI solutions,” said Secretary S. Krishnan.
Avataar will also publish a technical report outlining Varya’s model architecture, distillation methodology and benchmarks.
Distilled video generation is a model compression technique from machine learning where a compact “student” model replicates the outputs of a larger, slower “teacher” model transferring capabilities while eliminating redundant computation.
(IANS)
