Bikupan style

Hypernetwork trained on 600 unique images (1 200 after mirror) with a focus on faces (portraits). Low res images were upscaled, all images were stretched to 512x512, tagging done by "WD1.4 tagger" automatic1111 extension.

Recommend negative tags: author name, signature.

Low CFG (2.5-5.0) seems to produce results more consistently, extensive tagging recommended. Hypernetwork strength can be at least as low as 0.3 and still reproduce the style.

Model: AnythingV3

Hypernetwork structure: 1, 2, 2, 1

No dropout, no normalization

LN and steps: 5e-6 first 13 000 steps (focus on faces), remaining 37 000 steps 1e-5 (variety of poses, multiple subjects, etc)