Overview
Worked as a research intern at UNIST Embedded AI Lab, reviewing DDPM papers, training a toy diffusion model on MNIST, and exploring sampling optimization directions.
Technologies
Key Achievements
- Reviewed the DDPM (Denoising Diffusion Probabilistic Models) paper and organized key ideas from an implementation perspective
- Ran a hands-on toy project training a diffusion model on the MNIST dataset
- Extended the project with DDIM sampling and classifier-free guidance (CFG) to generate digit images conditioned on input numbers
- Investigated timestep sampling optimization studies; no significant gains were observed, but it became a valuable first image-generation model project experience