Model Description

0.1 - Preview

Attempt to sloppify Mugen SDXL, was trained with 3.7k images actually handpicked (unlike other liars) for 150 epochs on 1536 target resolution.

(this is a preview version, therefore it's allowed to be bad)

All samples are RAW GENS

things to be aware of before complaining:

• default gens are kinda meh, pupils started to appear properly, but it has a long way

• hands, 30% chance to fail (mainly finger count), depends on artist.

• specific gens on 1536 resolution will get stretched due to not being present in the dataset, happens very rarely.

• USE ARTISTS

Usage + what I used in the sample gens:

go to Mugen and make sure you can run it

• shift = 8.0

• scheduler = sgm_uniform

• sampler = euler_a2/euler_a/dpmpp_3m_sde_gpu/heun

• steps = 25

• cfg = 8.0 (can go wild here)

• resolutions (works both ways portrait/landscape):
  1536x1536 (1:1), 1344x1792 (3:4), 1280x1856 (2:3), 1152x2048 (9:16), 1024x2368 (9:21)
  1024 (works the same)

• nice artists/themes that you can prompt = rella/horiguchi yukiko/murota yuuhei/catbell/anime coloring

0.3 - Preview

everything improved, more pupils, you tell me, use artists

0.5 - Preview (coming soon)

added more artists to the mix, biasing eyes, keeping artist tokens to force model to override bad old art, added repeats for under-represented artists, going for double training amount