Cred founder and CEO Kunal Shah wasn’t exaggerating when he said India was now ready to build products for the world.
“Earlier, I used to think we were not really ready for that, but I feel now there is enough evidence that we might be able to create something interesting for the world,” said Shah in a recent interview.
Recently, Jivi’s purpose-built medical LLM Jivi MedX ranked number 1 on the Open Medical-LLM Leaderboard, outperforming OpenAI GPT-4 and Google’s Med-PaLM 2.
Meanwhile, Indian founders started QX Lab AI, which recently launched the Hybrid GenAI Multimodal Platform ‘Ask QX PRO’. Similarly, Quizizz supports millions of students in over 150 countries.
The list goes on.
Most recently, two Indian engineering students, Rudransh Agnihotri and Manasvi Kapoor, launched Mayakriti, an image generation platform that uses advanced GenAI to create lifelike visuals — from photorealistic portraits and personalised creations as well as in a variety of art styles such as cartoons, anime, and abstract art.
Rudransh, a third-year mechatronics engineering student from Delhi Skill and Entrepreneurship University, and Manasvi, currently in his second year, pursuing electronics and communication engineering, with a specialisation in AI and ML, from Netaji Subhas University of Technology founded FuturixAI and Quantum Works, a startup focusing on AI research and innovative solutions.
Mayakriti is a ‘Made in India’ product that utilises research, including papers such as Git Re-Basin and Arcee’s MergeKit, and concepts from mathematics and physics to create high-quality images without requiring the same amount of resources and computing power as needed by other popular models.
“To train something like GPT-4 in India is not possible in the foreseeable future due to the limitations of compute and training parameters. Even models like DALL·E 3, Midjourney, and Imagen require enormous training parameters. Since we are a research lab, we worked on the math and used techniques like SLERP in diffusion modelling to develop Mayakriti,” said the founders in an exclusive interview with AIM.
SLERP, or Spherical Linear Interpolation, is used to interpolate between the parameters of two models in a spherical space, ensuring smooth transitions and effective merging of models.
They highlighted that this technique has been used in video games and graphic rendering for some time now, but recent research showed its use in blending a model’s parameters into another so that the properties of both models remain.
“This is how we are able to give image generation qualities that are better than existing models and are not limited by the compute resources in India,” said Agnihotri, who believes that Indians have a higher than average math intellect and should focus on research to build models that can perform the same as the bigger models in fewer parameters.
AdapterFusion, which integrates multiple task-specific adapters into a base model, retaining its general capabilities while improving performance on specific tasks through lightweight modifications, is another technique that was used.
Anuvadini CEO Chandrasekhar Buddha said that the company, which is a Section 8 non-profit company under the Ministry of Education, recently provided FuturixAI and Quantum Works with NVIDIA A100 GPUs on Microsoft Azure Cloud network. Currently, Mayakriti is using 8 NVIDIA A100 GPUs (80 GB each) for deployment.
“I want to let people know through Mayakriti that if we are making that possible in just a few A100 GPUs by research and utilising core concepts of physics and math, then other Indians can also develop models as good as the big players by just focusing on the fundamentals,” said Agnihotri.
Image Generation Using Mayakriti
Apart from using existing open-source datasets, the founders also developed their own datasets. “In our colleges, our friends are graphic designers, so we asked them to go over the internet, collect some images and label them according to the prompt. This way, we collected up to image data from 3,000 images, and then multiplied that to the base variant,” they explained.
Mayakriti sets itself apart with its focus on hyper-realistic outputs and a wide range of customisable art styles. It employs several separate models and each model has its own speciality, from generating anime to creative arts.
Apart from letting users write their prompts, the platform also lets them choose resolution, environment, lighting, camera settings, composition, and style for the images.
It also comes with an in-built image editor, which the founders plan to improve by adding features, including the option to add text, which can allow users to create instant posters.
Agnihotri was once a JEE aspirant who couldn’t make it to an IIT, but that setback didn’t reduce his love for math. Now, with FuturixAI and Quantum Works, the young founder aims to push forward research in the AI field in India making use of research in math and physics.
“Google has the capability, dataset and compute, but at the same time, we have our own methods that are evolving with time,” he said.
In an attempt to make Mayakriti better than the image generation offered by Google and OpenAI, the founders plan to make the product free in the coming months for people to try globally and provide feedback.
Further, in the near future, the startup has another interesting product in the pipeline. They aim to release an AI platform that will let users train their ML models by uploading just a single dataset.
This is just the beginning and, soon in the future, we’ll see more such startups building from India, for the world!
