Suisui (Wuthering Waves) LoRA

Trigger word: sui
Trained base: waiIllustriousSDXL_v170 (Illustrious XL family)
Network: dim=32, alpha=16
Optimizer: Prodigy (lr=1.0)

Recommended settings

• Weight: 0.8 (identity metrics measured at 1.0; human quality calibration caps the recommendation at 0.8). At weight 1.0 complex / dynamic poses can smear — keep it around 0.75-0.8.

• Hi-res fix: recommended (~1.8x, denoise ~0.45) for clean lines on full-body / dynamic shots.

• Checkpoint: ep8 — passed the 5-criteria acceptance harness (identity / no-overtrain / NSFW / outfit).

Usage

1girl, solo, sui, blonde hair, blue inner hair, long hair, yellow eyes, tacet mark,
chinese clothes, white dress, bare shoulders, black gloves, folding fan,
<pose here>, <background here>,
masterpiece, best quality, amazing quality, absurdres, newest

Recommended weight 0.8. Trigger sui.

-- Identity (baked into the trigger) -- blonde long hair with a blue inner streak, yellow eyes, and the forehead Tacet Mark. These are baked in — keep hair/eye tags light and let the trigger carry the identity.

-- Outfit (add for the canonical look) -- chinese clothes / qipao in white・blue・gold, bare shoulders, black gloves, hair ornament, folding fan.

-- Style note (IMPORTANT) -- For clean anime shading (not 3D / semi-real skin), put anti-realistic terms in the NEGATIVE prompt: realistic, 3d, photorealistic, shiny skin, oiled body, semi-realistic, render. The negatives alone are enough — you do NOT need 2D style tags in the positive. This matters most on nude / exposed skin.

-- NSFW -- Identity holds when undressed / nude; the Tacet Mark persists through explicit scenes. Keep weight ~0.8 for coherent full-body / dynamic ero.

Tested checkpoints

• ✅ waiIllustriousSDXL_v170 (trained base, identity 100%)

• 🔍 Other Illustrious XL fine-tunes (untested; start weight 0.6-0.8)

• ❌ Pony / SDXL base (architecture mismatch, untested)

Caveats

• On checkpoints other than the training base, the recommended weight shifts (try 0.7-1.2).

• Only outfit variations present in the dataset reproduce strongly.

• Face / hair color / eye color are baked into the trigger — no need to over-specify them (reproduction is higher when you don't).