Prompt:

A modern smartphone picture of a man riding a motorcycle in front of a row of brightly-colored buildings.

Settings:

Rendered using various steps and CFG values, Euler a for the sampler, no manual VAE override (default VAE), and no refiner model. All images were generated at 1024*1024. This is using the 1.0 version of SDXL.

Summary:

Subjectively, 50-200 steps look best, with higher step counts generally adding more detail. A CFG of 7-10 is generally best, as going over will tend to overbake, as we’ve seen in earlier SD models. Prompting and the refiner model aside, it seems like the fundamental settings you’re used to using will probably still hold true for SDXL. Granted, prompting is a bit easier for photorealistic outputs now, and the refiner model might allow you to use fewer steps for the initial generation with the base model.