📒 Read this carefully if you are using Pony Diffusion V6 XL

Anime characters are often depicted with exaggerated eyes, which can look unsettling when applied to real-life features. The Pony base model (Diffusion V6 XL) is trained using anime-style source files, meaning it struggles to generate realistic or 2.5D-style images. After 300 trials, I have found a way to avoid the weird eye glitches.

📒 Add this into positive prompt in addition to trigger word (for V1.0 only):

perfect eyes, detailed face

📒 Add this into negative prompt (for V1.0 only):

eye shadow, eyeliner, eye line, poor drawn eyes, child, source_pony, source_furry, source_anime,

📒 Settings for A1111:

lora weight 0.75

CFG 5.5-7.5

Steps 22-50

📒 Update Note:

V1: num repeats of 8 / epoch too low / expressive eyes issue

V1.1: num repeats of 15 / too many repeats / over training

V1.2: num repeats of 10 / too many repeats / over training

V1.3: num repeats of 3 / not enough repeats / under training

V1.4: num repeats of 5 / ?

🎉 This lora is a branch spun off from an experimental lora complex from 2024. It's part of my goal to create 100 lora, and I feel it's good enough to stand independently.

📒 Checkpoint Rating: