From imaging algorithms to quantum methods Seminar

Comparing GAN and Diffusion Models on an Interpretable 3D Voxel Dataset

24 Nov 2025, 10:00

1h

https://cern.zoom.us/j/66151941204?pwd=n7upvvZYibexBhbtyn5kvTpy36L0Wo.1 (Zoom)

Speaker

Aleksander Ogonowski

Description

Overview of generative models evaluated on a simple, visually interpretable 3D voxel dataset:

• Creation of an easy-to-interpret 3D voxel dataset derived from 2D characters
• Conditional WGAN-GP for fast 3D volume generation
• Conditional diffusion model for smoother and more coherent volumetric synthesis
• Evaluation using KID, PCC, and Separation Power metrics
• Comparative analysis of generative behavior across increasing dataset complexity