What is the usual number of steps needed to generate decent images?

#41
by Dinghuai - opened

In the document the num_inference_steps is set to 28. I wonder where does this number come from?

Does anyone know if we can use less steps to generate decent-looking images?

@Dinghuai Depends on what you mean by decent.
Sometimes, 20 steps works fine enough so you can try that. In my opinion, usually the best results come from around 30 or 35 steps where you get good text and quality. Still doesn't really solve human issues but overall good.

Sign up or log in to comment