For best outputs:

Version 3.0:

Sampler: DPM++ 2M Karras / DPM++ 2M SDE Karras.

Steps: 28 ~ 36, cfg 5 ~ 6.5, CLIP skip 2.

VAE: none required.

Pos prompt: nothing remarkable, just what you want to see. Version 3.0 listens better than 2.0, usually not big weights required.

Neg prompt: can be left blank or just what you dont want to see, for example, "render, cgi, drawing, cartoon, painting, illustration, bad quality, low res, watermark, logo, bad quality, low res, boring, blurry, ugly, clean, cute, thin, child, kid, nsfw".

Version 2.0:

Sampler: DPM++ 2M Karras.

Steps: 30, cfg 7, CLIP skip 2.

VAE: blessed2_fp16.safetensors.

Pos prompt: use the LoRa "add_detail" with a weight of 1 at the end of your pos prompt (found in my 2.0 version runs).

Neg prompt: "render, cgi, drawing, cartoon, painting, illustration, bad quality, low res, watermark, logo, bad quality, low res, boring, blurry, ugly, clean, cute, thin".

What changes from v2.0 to v3.0 Trinity: input_image feature, overall quality and realism greatly increased.

Checkpoint Merge.