This project uses a Low-Rank Approximation of a pretrained Stable Diffusion model to generate beautiful scenery. Examples of such pictures are shown below, and you can use the model generator freely at this link. Note: To maintain low compute costs, the website rate limits users to 100 queries per day.
Front-end:
- React
- Material UI
Back-end:
- Express
Cloud:
- AWS EC2
- AWS Lambda
- NGrok
Model Development:
- PyTorch
- Huggingface
- Weights & Biases
- Beautiful Soup
Low Rank Approximation is a technique which lossy compresses Singular Value Decomposition matrices into lower rank while maintaining the “energy” within the matrix, given by the monotonically decreasing sequence of singular values. In a process called Low Rank Adaptation, we freeze the pretrained model weights of a model and inject trainable Low Rank decomposition matrices onto their weights to quickly adapt them to specific tasks.
This Low Rank Adaptation uses the base model, Anything V5 and finetunes it using a repository of 171 images of anime scenery, which you can find here
