This model was trained on meticulously selected datasets over a long period. It's able to produce very realistic images with intricate and realistic backgrounds. I've included a number of sample photos that were all generated at 40 steps, no LoRAs, no embeddings, no negative prompts.
I believe this model shines the best at 60-80 steps, with various embedding, refiner checkpoints, and it plays nice with LoRA's, so you should be able to extend the model to do very interesting things.
I find it's core strength to be prompt adherence at low tokens and working off of simple prompts. I'll put some recommendations at the bottom, but I have got great results across different samplers, though Eula samples do grade more cartoonish, which might be your thing.
Recommended:
Sampler: DPM++ 2M SDE (Or something in that family.)
Scheduler: Karras
Steps: 60-80
Resolution: 512x768
Upscaler: ESGRAN 4x (Sample pictures are lanczos)