Character LoRA for Anima (anima_baseV10) — Illustrious (光辉) from Azur Lane 碧蓝航线.
What it does
Generates the Azur Lane shipgirl Illustrious / 光辉 — silver-white long hair, blue eyes, blue flower hair ornament, mole under the eye. The character identity is fully baked into the trigger word, so you only need to write the outfit and the scene.
Trigger word
ghui01Minimum working prompt:
masterpiece, best quality, score_9, safe, ghui01, 1girl, solo, facing viewer, looking at viewerHow to use
Always keep
facing viewer, looking at viewer— without them the model may randomly turn the character away from the camera
Do NOT write hair / eye / hair-ornament tags — they are already learned
Write any concrete outfit tags and she dresses accordingly: white dress, wedding dress, kimono, christmas sweater, coat and scarf, hoodie, cozy sweater — all verified
Without outfit tags she defaults to a white camisole style
Without composition tags the output is a close-up bust shot
Recommended settings
LoRA strength: 1.0 (0.8 for more prompt-obedient outfits, 1.2 for stronger character traits)
Steps: 30 · CFG: 4.5 · Sampler: euler + simple
Resolution: 832×1216
Hi-res fix: latent upscale 1.5 → 1248×1824, 20 steps, denoise 0.42
Example prompt
masterpiece, best quality, score_9, safe, ghui01, 1girl, solo,
white dress, white thighhighs, white gloves,
standing, facing viewer, looking at viewer, gentle smile,
cowboy shot, soft lighting, highly detailed illustrationNegative prompt:
worst quality, low quality, score_1, score_2, score_3, blurry, jpeg artifacts,
watermark, signature, artist name, fake text, japanese text,
bad hands, extra fingers, fused fingers, 2girls, multiple girls,
nude, naked, nsfw, nipples, areola, from behind, back to viewer,
3d, cel shading, renderStrengths
Face consistency — tested on 9 random seeds: 9/9 correct character, 8/9 clean front-facing face
Outfit flexibility — concrete garment tags (material + style) are followed accurately
Soft-light portraits — warm indoor light, morning window light, lantern glow, autumn bokeh all work well
Limitations
Towel / wrapped-around-body garments are not understood (
towel wrappedtends to render a loose shirt)
Full-body beach shots may show stray floating fabric near shoulders or hips
Abstract outfit words (e.g. only
dress) fall back to the training-set default style
If clothing that should be loose renders skin-tight, add
wide sleeves, loose fabric, flowing hemto the positive andtight clothing, bodysuit, skin tightto the negative
Do NOT add
huge breaststo the negative prompt — the body shape is an intentional character trait
Training details
Base model: Anima (anima_baseV10)
Dataset: 42 images, 1024px, WD14 auto-tagged with identity tags removed, no artist tags
Network: LoRA dim 32 / alpha 16, UNET only, bf16
Optimizer: AdamW8bit, lr 5e-5, cosine_with_restarts
Schedule: timestep sampling shift, discrete_flow_shift 3.0
Training: 15 epochs / 6300 steps — converged by epoch 8
Disclaimer
Fan-made character LoRA trained on fan art. The character Illustrious / 光辉 is © Manjuu / Yongshi (Azur Lane). Non-commercial use only. Do not use this model to generate illegal or harmful content.
Description
Initial release.
- Base model: Anima (anima_baseV10)
- 42 images, dim 32 / alpha 16, 15 epochs (6300 steps)
- Converged by epoch 8; no further checkpoints provided on purpose
- Trigger: ghui01
- Strength 1.0 recommended
Tested: 9 random seeds → 9/9 correct character.
Known limits: towel/wrapped garments, full-body beach shots, abstract outfit words.




