Artificial Intelligence in Video Production – How Varya is 20x Less Expensive

I have always checked the counter when I am generating some video snippets for my blog posts or courses. AI video generation is still very expensive. Which is why I was intrigued to see this: Indian startup Avataar AI (funded by Peak XV) has launched their new video generation model called Varya which costs ₹0.48 (half a US cent) per second of generated video. This is compared to $0.10 / sec (and higher) for Google’s Veo, Kling, Luma and Runway. So Varya is ~20x cheaper to do the same thing i.e. write a prompt or give an image and get back a video.

And performance is equally impressive. Varya can generate 5 seconds of 720p video in 45 seconds using a single Nvidia H200 GPU. Meanwhile, the original model that Varya is distilled from (Alibaba’s open Wan 2.2) takes ~1,230 seconds to generate the same result. So that’s ~10x faster since Varya uses 4 generative steps compared to Wan’s 50.

1Distillation is eating the price list

But what excites me is the fact that Avataar did not build a cutting edge model from scratch. They used a pre-trained open source model and distilled it to create a small, fast version that retains most of the quality but at a much lower cost. This is exactly what happened with chatbots — they got distilled and became affordable. Now the same thing is happening with video generation. Avataar is one of the 12 companies participating in the India AI Mission (a $1.2bn government initiative to subsidize compute in return for open sourcing the models) and hence they have made Varya’s weights available on the AI Kosh portal and have opened up a public beta on their website. And they have intentionally trained their model on Indian festivals, cuisine, attire and architecture so that they do not generate stereotypical brochure material. As Rajan Anandan (Peak XV) said “Lowering costs is the single biggest factor that will drive AI adoption in India”.

So what does this mean for you if you are not based in India? Well, sooner or later these costs will come down globally. We have seen this happen with chat subscription services where they started in the low cost segments in developing countries before coming to everyone else. Today video generation is the most expensive thing you can generate with AI and that’s why it will be the next to fall to the power of distillation. If it costs half a cent to generate a 720p video, then video generation will become a commodity and every app you use will have a video generator button right there. Whether it’s for presentations, marketplaces, family photo albums etc.

Verdict: In the world of AI video generation, the battle is no longer just about who can render the prettiest pixels, but also who can do it a hundred times cheaper. And with distillation we just saw how the price can be cut by 20x in one shot. Soon enough we will have pretty pixels too.