For the best outputs:

Sampler: DPM++ 2M Karras.

Steps: 58, cfg 7 ~ 9, CLIP skip 2.

VAE: clearvae_v2.3_fp16.safetensors / kl-f8-anime2_fp16.safetensors

Pos prompt: just what you want to see. To reach max realism, for example, "epiC35mm, Ultra Realistic concept photo of" + your text + "cinematic, high budget, epic realistic, photorealistic, highest quality, skin texture, intricate details, (cinematic lighting), RAW photo, 8k, , ", being the LoRa's "epiC35mm" and "more_details" included in my v1.5 example runs.

Neg prompt: "airbrush, photoshop, plastic doll, plastic skin, easynegative, monochrome, (low quality, worst quality:1.4), illustration, cg, 3d, render, anime", being the embedding "easynegative" included in my example runs for both versions.

What changes from v3.0 to v4.0: input_image feature.

Checkpoint Merge.