Understanding generative models and CLIP

• 

  Generated with DDPMpipleline which was pre-trained on a celebrity dataset

• 

  With only a U-Net, we can not generate good quality images, we should use DDPMschdulers.

• 

  Generative models do not perform well on generative faces, even if they have been trained on those faces.

• 

  Generating images with U-Net and schedulers result in a much better generated image.

• 

  Generated with DDPMpipleline which was pre-trained on a celebrity dataset

• 

  Also generated with DDPMpipleline, pre-trained on a celebrity dataset

• 

• 

Generative models are now everywhere. We can try them online, but how do they learn different structures?

Built With

• python
• pytorch