For best outputs:
Sampler: DPM++ 2M Karras.
Steps: 25, cfg 7, CLIP skip 2.
VAE: kl-f8-anime2_fp16.safetensors / clearvae_v2.3_fp16.safetensors if you want less saturated images.
Pos prompt: just what you want to see. Usually the model listens well and requires of no big weights.
Neg prompt: "ng_deepnegative_v1_75t, negative_hand-neg, (worst quality, low quality:2), lowres, bad anatomy, normal quality, monochrome, grayscale, text, signature, watermark, logo, loli, teen, child, aged down, lolita, shota, minor, underage, flat chest", being the "ng_deepnegative_v1_75t" and "negative_hand-neg" embeddings included in my example runs.
What changes from v2.0 to v2.5: input_image feature.
Checkpoint Merge.
.webp)
.webp)
.webp)
.webp)
For best outputs:
Sampler: DPM++ 2M Karras.
Steps: 25, cfg 7, CLIP skip 2.
VAE: kl-f8-anime2_fp16.safetensors / clearvae_v2.3_fp16.safetensors if you want less saturated images.
Pos prompt: just what you want to see. Usually the model listens well and requires of no big weights.
Neg prompt: "ng_deepnegative_v1_75t, negative_hand-neg, (worst quality, low quality:2), lowres, bad anatomy, normal quality, monochrome, grayscale, text, signature, watermark, logo, loli, teen, child, aged down, lolita, shota, minor, underage, flat chest", being the "ng_deepnegative_v1_75t" and "negative_hand-neg" embeddings included in my example runs.
What changes from v2.0 to v2.5: input_image feature.
Checkpoint Merge.
