helblazer811/Diffusion-Explorer
Interactive visualizations of the geometric intuition behind diffusion models.
What it solves
Diffusion Explorer is an educational tool designed to help users understand the geometric intuitions behind diffusion and flow-based generative models. It bridges the gap between complex mathematical concepts and visual intuition by providing interactive visualizations of how these models generate data.
How it works
The tool implements several training objectives, including Flow Matching and Denoising Score Matching. It allows users to visualize the dynamics of generated samples over time for pretrained models and observe how samples evolve during the training process. Additionally, it supports training on custom, hand-drawn distributions to make the learning process more tangible.
Who it’s for
This tool is designed for students, researchers, and developers who want to visually explore the concepts of diffusion and flow-based generative AI, specifically those interested in learning about Rectified Flow and optimal transport.
Highlights
- Interactive visualizations of generative model dynamics.
- Support for Flow Matching and Denoising Score Matching objectives.
- Ability to train models on custom hand-drawn distributions.
- Dedicated explainer for Rectified Flow to demonstrate the connection between straight paths and optimal transport.