For best outputs:

Version 3.0:

Sampler: DPM++ 2M Karras.

Steps: 18, cfg 2.5, CLIP skip 2.

VAE: vae-ft-mse-840000-ema-pruned_fp16.safetensors (optional for a little bit more of image's detail).

Pos prompt: add the LoRa "add_detail" with a weight of 1 (found in my v3.0 example runs).

Neg prompt: "(EasyNegative, cartoon, painting, illustration, (grayscale:1.4), normal quality:1.8, 3D rendering, ugly, fat, frame", being the "EasyNegative" embedding found in my 3.0 example runs.

Version 2.50:

Sampler: DPM++ 2M SDE Karras.

Steps: 50, cfg 7, CLIP skip 2.

VAE: vae-ft-mse-840000-ema-pruned_fp16.safetensors.

Pos prompt: add the LoRa "more_details" with a weight of 0.8 at the end of your pos (found in my 2.5 version runs).

Neg prompt: "bad-hands-5, By bad artist -neg, bad-image-v2-39000", being such embeddings found in my 2.5 version runs.

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

Checkpoint Merge.