Inspiration
Upon meeting each other, we realized our shared love in computer vision and the deep and vastly dazzling space. We wanted to create the feeling of floating the deep and unending expanse of space. We also wanted to create an image generation model to create realistic images of space. By searching high and low for a curated deep space dataset (about 2.7k images) with each image containing vase about of data.
What it does
By opening the webpage, you can click on any of the celestial objects on the screen(the star dust nebula, planets, stars, supergalaxy, blackholes) would give you generated images related to the celestial object that are grounded in reality
How we built it
Challenges we ran into
frontend - image 1: initial our webpage was barren not giving the feeling of looking at the bright stars or the colorful nebula that cover this universe. It feel more like buttons that give out the generated images. image 2: trying to implement the cloud feeling about the nebula kinda fail here as this nebula button just covers the entire center. image 3: we lost the starry background for the favor of highlighting the nebula but we pushed through found the cloud like state we wanted in our end product
backend: integrating the frontend with the backend was very difficult. It took us a number of setbacks and issues before we were able to get it fixed. also, when clicking on a space object image, generating an image did not work for a while. to fix it, we had to create a python server file to properly connect to modal as well. Procuring the deep space dataset was different problem as finding any annotated data on this was kinda niche, and due to the nature of the individual image, them been massive, breaking it down and make it useable to train the model was a different hurdle we had to cross
Accomplishments that we're proud of
Frontend - the overall design, the cloud like moments of the star dust nebula Backend - training the diffusion model and integrating it with the frontend
What we learned
Through this project, we learned how to fine-tune a large image generation model (Stable Diffusion XL) on a custom dataset using Dreambooth LoRA, and how to deploy GPU-intensive machine learning workloads serverlessly using Modal.
What's next for Deep Space Diffusion
Next, we'd love to expand the planetary system significantly. We could classify planets as rocky, gas giant, ice world, or ocean world, and have the model generate scientifically plausible imagery for each.
Log in or sign up for Devpost to join the conversation.