Semantic Image Synthesis with EDM & eDiff-I

Model Archirecture

Results

All the testing images are not in the training set.

EDM

128 x 128 images , 1000 epochs

Reference

• Learning Transferable Visual Models From Natural Language Supervision
• Elucidating the Design Space of Diffusion-Based Generative Models
• eDiff-I: Text-to-Image Diffusion Models with Ensemble of Expert Denoisers