For best outputs and usage:

Sampler: DPM++ 2M Karras.

Steps: 36, cfg 9, CLIP skip 2.

VAE: sdxl_vae.safetensors

Usage: the model listens well and usually doesn't need of big weights. The only big weights used in my example runs were in order to push the model's realism to its max. By default, this model is set to something in between semi-realism and realism, but it can be pushed to realism by doing the next:

Pos prompt: wrap your text between "score_8_up, score_7_up, score_6_up, score_5_up, (overview:1.5), realism, (realistic:1.8), high quality, Kodak, Analog style, Leica, Fujiflim, cinematic, detailed skin texture, soft light, sharp, best quality, masterpiece, realistic, hyperdetailed, hyperrealistic, photorealistic, 8k, high quality, absurd res, natural lighting, intricate, soft warm lighting" and "real photo, skin texture, skin imperfections, skin pores, sharp details, 8k, HDR, f.1/2, 70mm lens".

Neg prompt: use this one "score_6, score_5, score_4, smile, (teeth), clenching teeth, low res, poorly drawn, low quality, jpeg artifacts, film grain, old, old woman, mature woman, worst quality, low quality, normal quality, messy drawing, amateur drawing, lowres, bad anatomy, bad hands, source_furry, source_pony, source_cartoon, comic, child, loli, kid, underage, source filmmaker, 3d, censor, bar censor, mosaic censorship, epiCNegative", being the "epiCNegative" embedding included in my example runs.

Recommended 1024 x 1024 size.