V2 — retrained on NoobAI-XL, tag-based captioning

V1 was trained on Illustrious with natural-language captions, which is the wrong

labeling scheme for the Illustrious/NoobAI family — the character token never bound

properly and facial likeness came out soft. V2 fixes that:

- Base model: NoobAI-XL (EPS)

- Danbooru tag captioning instead of prose captions

- Cleaned dataset — inconsistent and low-face-resolution samples removed

Trigger word: chris_ml

Recommended usage

- Works best on NoobAI-XL based checkpoints; Illustrious checkpoints (WAI, etc.)

also work thanks to the shared lineage — drop strength slightly if the face

drifts.

- LoRA strength: 0.8–1.0

- Sampler: Euler a / DPM++ 2M, CFG 5–7, 28–30 steps

- A face detailer / ADetailer pass at ~0.4–0.5 denoise noticeably sharpens likeness

at full-body framings.

- Keep the trigger word at the very front of the prompt.

Prompt example

`urasmanga, 1boy, solo, white shirt, sitting at a cafe table, looking at viewer,

warm sunlight, indoors, upper body, masterpiece, best quality`

Stylistic tags (muscular, bara, beard length, tattoos, hair color) stay fully

promptable — they were tagged in the dataset, so you can steer them freely without

breaking the character.