Description

This is a WAI / Illustrious-SDXL character LoRA for Shuidengxin - Whisper Into Dream / 水灯心「轻语入梦」, trained with the trigger shuidengxin.

It is designed to recall:

• very long blonde hair with wispy bangs and ahoge

• large pink hair bow and pink headband

• blue eyes and fair skin

• floppy dog ears and large fluffy brown tail

• white blouse with puffy sleeves and frilled collar

• blue neck ribbon, black chest bow and daisy ornament

• pastel pink / lavender layered dress with apron-like skirt

• mismatched legwear and black mary jane shoes

The LoRA is strongest for character identity, hairstyle, outfit structure and the main visual motif of this skin/outfit. It was trained and evaluated primarily for WAI v17 / Illustrious-style SDXL checkpoints. It is not intended as a Pony, realistic XL or non-Illustrious compatibility release.

Training metadata records SFW samples: unknown, NSFW samples: unknown. This does not mean the preview images are NSFW.

Recommended Usage

Default Weight

0.85

This is the recommended balance between identity recall, outfit stability and prompt obedience.

Useful Range

0.75-0.9

Motion / Composition First

0.6 - 0.75

Use a lower weight when pose, motion, camera angle or background is more important than maximum outfit fidelity.

Avoid as Default

0.95 - 1.0

Very high weights may over-strengthen outfit details, reduce pose freedom, or pull the image toward repeated compositions.

Basic Prompt

<lora:shuidengxin_recommended_e000009:0.85>, shuidengxin, 1girl, solo, no text, looking at viewer, upper body, alluring smile, simple background

Full Character Prompt

<lora:shuidengxin_recommended_e000009:0.85>, masterpiece, best quality, amazing quality, very aesthetic, absurdres, shuidengxin, 1girl, solo, very long blonde hair, floor-length hair, thick hair, long sidelocks, wispy bangs, ahoge, large pink hair bow, pink headband, blue eyes, fair skin, floppy dog ears, brown animal ears, large fluffy brown tail, dog girl, white blouse, puffy sleeves, frilled collar, blue neck ribbon, black chest bow, daisy ornament, pastel pink dress, lavender dress, layered dress, apron dress, asymmetrical skirt, scalloped hem, blue inner skirt, ribbons, frills, mismatched legwear, single thighhigh, one bare leg, pink argyle thighhigh, lavender frilled ankle socks, black mary janes, glossy black shoes, round-toe shoes, ankle strap shoes, no text, detailed outfit, soft lighting

Negative Prompt

low quality, worst quality, lowres, bad anatomy, bad hands, bad fingers, extra fingers, missing fingers, extra limbs, disfigured, deformed, blurry, jpeg artifacts, text, watermark, logo, signature, arknights logo, chinese text, english text, caption, artist name, bad face, bad eyes, multiple girls, 2girls, cropped head, out of frame

Example Prompts

Portrait

<lora:shuidengxin_recommended_e000009:0.8>, shuidengxin, 1girl, solo, no text, portrait, close-up, looking at viewer, beautiful face, detailed eyes, simple background

Full Body

<lora:shuidengxin_recommended_e000009:0.85>, shuidengxin, 1girl, solo, no text, full body, standing, elegant pose, clear outfit, clear footwear, simple background

Seated / Sensual Showcase

<lora:shuidengxin_recommended_e000009:0.85>, shuidengxin, 1girl, solo, no text, sitting, crossed legs, leaning back, charming smile, looking at viewer, soft warm lighting, elegant lounge

Recommended Generation Settings

These settings were used for the included showcase images:

Checkpoint: WAI-Nsfw-Illustrious-17
Resolution: 1024 x 1024
Sampler: DPM++ 2M
Scheduler: Karras
Steps: 30
CFG scale: 6.5
LoRA weight: 0.85

All showcase prompts include no text; the negative prompt includes text, watermark, logo and signature to suppress Civitai preview text pollution.

Known Limitations

• The LoRA was trained for WAI / Illustrious-SDXL. Other model families may change the identity and outfit details.

• Higher weights can improve identity recall, but may reduce pose and background flexibility.

• Very complex poses may require lower LoRA weight and stronger action tags.

• Combining this LoRA with strong outfit or style LoRAs may cause outfit conflicts.

• If unwanted text or logos appear, keep no text in the positive prompt and keep text, watermark, logo, signature in the negative prompt.

Training Summary

Training images: 100 PNG
Caption files: 100 TXT
SFW samples: unknown
NSFW samples: unknown
Base model: WAI-Nsfw-Illustrious-17
Training type: SDXL LoRA, U-Net only
Network dim / alpha: 32 / 16
Network dropout: 0.05
Optimizer: AdamW8bit
Precision: bf16
Recommended checkpoint: epoch 9
Recommended file: shuidengxin_recommended_e000009.safetensors
Recommended weight: 0.85

The selected checkpoint was chosen after Forge visual evaluation across multiple prompts and LoRA weights. The final package keeps only the recommended checkpoint for public release.