I’m working on a prototype application, and for my use-case DeepFloyd IF gives me the best results by far. I was using replicate.ai, but the reliability and cold start times are unacceptable for my purposes (and DeepFloyd IF cannot be used in commercial products).

I think what made IF so good for my use-case is that it starts by generating a very small image. I need images generated with a single subject taking up most of the image and minimal background. I think IF is biased to these kinds of images since it starts with a very small image.

Currently using Dall-E 2, which is ok, but not nearly as good as IF. Stable Diffusion sometimes produces very weird images (haven’t tried SD XL yet).