For best outputs:

Sampler: DPM++ 2M Karras.

Steps: 35, cfg 8, CLIP skip 2.

VAE:

For realistic barbies: clearvae_v2.3_fp16.safetensors / blessed2_fp16.safetensors

For doll-like barbies: kl-f8-anime2_fp16.safetensors

Pos prompt:

For realistic barbies: use realism-like words, like "photorealistic", ect.

For doll-like barbies: use keywords like "3D, pixar style" ect. Also add the keywords "doll" and "plastic like", it will add joints like a doll.

For both styles: add the LoRa's "FaceBeauty_qinglong_V3" (weight 0.9) and "more_details" (weight 0.7) at the end of your pos to get some extra face / overall detail (both included in my example runs).

Neg prompt:

For realistic barbies: adding the keywords "2d, cartoon" may help with a bit more of realism.

For doll-like barbies: adding the keywords "photorealism, realism, lifelike, 2d, cartoon" will increase the doll look.

For both styles: the embeddings "easynegative" and "bad-hands-5" do help some, you can find them in my example runs.

Overall prompts note: not using any kind of realism / barbie doll like keywords creates a mix of both.

What changes from v1.0 to v1.5: input_image feature.

Checkpoint Merge.