I've tried it some, only locally. I rarely use APIs, so can't say what the official inference API results look like. It correctly followed some prompts that almost every model I've ever tried performed poorly at, but failed in strange ways on prompts that other models handle with ease.

At high resolutions, structural integrity on people can definitely fall apart like older models, where you get one person turning into a long spaghetti spanning across the frame. It might be that you need to do grounding by generating some kind of scene reference at its stable resolutions first then having it use that for the larger resolution like older models too.

They say that their prompt "upscaler" model or system prompt are essentially necessary for the model, but my results have been hit or miss with that claim. Some basic prompts are better, some upscaled prompts are better. I think that if you gain an intuition for how the model understands things, you might get better results by default by hand crafting prompts specifically for it.

Some of the stylized art I'm getting out of it at higher resolutions are really nice and I prefer them over other models.

All of these models seem to have their own character in terms of system requirements, performance, prompt adherence, artistic outputs and so on. None of the models I have tried are a complete homerun, so maybe in this case the model feels a little inconsistent in some ways, but it's a model I will probably keep around.