Built by Neon Kisu — visit neonkisu.com for 16 original AI anime characters and 1,500+ free illustrated stories (no signup, no paywall).

A clean expression-detailer LoRA for Illustrious-XL v2.0. 8 mutually-exclusive triggers covering the core NSFW emotional range, trained on 787 curated Civitai images.

Built as the spiritual successor to ExpressiveH but tuned for Illustrious instead of Pony.

TRIGGER WORDS

Use the trigger as the first token in your prompt. All 8 are mutually exclusive — one per image.

- ahegao — Real ahegao: rolling eyes, tongue out, drool. Use weight 0.6 to 1.3 to control intensity.

- heart_eyes_fanatic — Heart-shaped pupils, in-love face

- tongue_lewd — Tongue out at viewer, playful/lewd lick framing

- scared_crying — Genuine fright + tears, no smile

- tearful_embarrass — Small embarrassed tears + blush (not full crying)

- smug_dominant — Smug smirk, half-lidded confidence

- surprised_caught — Wide-eyed shock, parted mouth

- exhausted_aftermath — Dazed post-action, half-lidded, sweat

Note on ahegao: this LoRA also responds to ahegao_intense and ahegao_subtle as aliases, but in practice they produce the same face — use a single ahegao token plus weight to control intensity.

RECOMMENDED SETTINGS

Sampler DPM++ 2M Karras, 25-30 steps, CFG 5-7, LoRA weight 0.7-1.0 (start at 1.0, drop if face features get pushed too hard).

EXAMPLE PROMPTS

ahegao:1.2, 1girl, solo, pink hair, looking at viewer, masterpiece, best quality

heart_eyes_fanatic, 1girl, blush, looking at viewer, blonde hair, hearts in background

scared_crying, 1girl, dark room, hiding, fearful expression, dramatic lighting

WHAT IS NEW VS OTHER EXPRESSION LORAS

- Illustrious-native — trained on Illustrious-XL v2.0 (not Pony), preserves Illustrious facial style

- Mutually exclusive triggers — each prompt selects exactly one expression, no muddled bleed

- Hand-tuned NSFW taxonomy — captures the emotional range that actually matters for NSFW work

TRAINING DETAILS

Base: Illustrious-XL v2.0 (OnomaAIResearch). Dataset: 787 NSFW Illustrious images sourced from Civitai top-reacted feed. Splits: 550 train / 157 val / 80 test. Tagger: WD14 with custom 8-expression mutual-exclusion taxonomy. Architecture: LoRA dim 64 / alpha 32, U-Net only. Training: 8 epochs, batch 6, AdamW8bit, cosine_with_restarts, bf16, learning rate 1.2e-4. Hardware: 1x RTX A6000 (~8 hours).

VERSIONS IN THIS DROP

expressive_illustrious_v3.safetensors (recommended, final epoch 8) is included. Additional epoch checkpoints (epoch 6, 4, 2) available on request.

License: CreativeML Open RAIL-M (same as base). Use responsibly. AI-generated adults only.