Varys is an interior design tool integrating Stable Diffusion and GPT that allows users to draw a floorplan, convert any part of the design into SD AI renders, and chat with a specialised GPT for design specs and help. I have collaborated closely with a team of 5 people on this project, the team consists of Lucas, who is an architect and interior designer, also the project initiator; me and Jeff as the software engineer; Melissa as the product manager and Liu as the algorithm engineer.
I'm responsible for the multi-language landing site for the product, which has gathered 2700+ waitlist users, most of whom are validated business opportunities with corporate email and titles. As the motion and UX designer, Jeff has made the amazing concept motion graphics shown above and motion components on the site. I have contributed around 70% of the site, including layouts, forms, animations and interactive components with Framer Motion. I18n and language switch are also incorporated into the site, with static export deployed on Cloudflare pages, ensuring speed globally.
I have also contributed to Stable Duffision & Control-Net integration, and GPT LangChain retrieval training. The SD is deployed to replicate, a machine learning cloud provider, to minimize the cost of running the GPU services. We use the Realistic Vision model, Lora and Control-Net to achieve the render of the user's 3D floorplan design, where we have integrated them into a custom image with Cog to be built and deployed on replicate.
The editor allows user to build their floor plan with common types of furniture, and then it is converted to a 3D environment in real time with threejs enabling the user to select a view to generate the render, you can read more about the amazing work on Jeff's blog.