Nanami Idol - [Anima DiT]

A decoupled character LoRA trained on CircleStoneLabs/Anima-Base-v1.0 using a curated 55-image dataset. Facial likeness, eye geometry, and anatomical proportions are completely isolated from clothing. The trigger word activates character identity without locking you into a pre-baked wardrobe.

---

Recommended Generation Parameters

Setting	Recommended Value	Notes
Base Model	CircleStoneLabs/Anima-Base-v1.0	Compatible with Anima DiT checkpoints
Trigger Word	Nanami-Character	Prepend to prompt
LoRA Weight	0.80	Sweet spot: 0.75 - 0.85 (0.60 for styling, 1.0 for likeness lock)
Sampling Steps	30	Range: 28 - 35
CFG Scale	4.0 - 4.5	Keep between 3.5 - 5.0; DiT burns above 5.5
Sampler	Euler / Euler a / er_sde	Euler a for softer lines; er_sde for crisp linework
Resolutions	832x1216 (Portrait) | 1024x1024 (Square)	768x1344 for full-body shots

---

Curated Prompt Presets

1. Starlight Idol Stage (Performance Center)

Prompt:
masterpiece, best quality, Nanami-Character, solo, 1girl, idol stage costume, frilled sleeveless bodice, matching layered ruff skirt, satin elbow gloves, mini top hat hairpin, holding microphone, dynamic singing pose, winking, stadium concert stage, volumetric stage lights, lens flare, bokeh, dramatic lighting

2. Techwear & Street Beat (Harajuku / Cyber)

Prompt:
masterpiece, best quality, Nanami-Character, solo, 1girl, techwear aesthetic, cropped windbreaker jacket, black sports top, high-waisted cargo joggers, multi-straps, chunky sneakers, fingerless gloves, leaning against graffiti wall, neon signs, rainy tokyo alley, wet reflections, vibrant contrast

3. Classic Academy (Serafuku Uniform)

Prompt:
masterpiece, best quality, Nanami-Character, solo, 1girl, traditional serafuku, white sailor shirt, navy blue sailor collar, red neckerchief, pleated navy skirt, dark knee socks, loafers, looking at viewer, gentle smile, sunlit classroom, wooden desks, soft morning sunlight, ray tracing

4. Pop Gothic / Dark Alt Aesthetic

Prompt:
masterpiece, best quality, score_7, Nanami-Character, solo, 1girl, pop gothic style, black velvet dress, white lace trim, leather harness choker, spiderweb lace tights, platform mary janes, twintails, subtle smirk, dimly lit antique room, crimson velvet curtains, baroque mirror, rim lighting

5. Casual Weekend (Oversized Street Knit)

Prompt:
masterpiece, best quality, Nanami-Character, solo, 1girl, oversized chunky knit turtleneck sweater, pleated beige tennis skirt, thigh-high socks, white sneakers, hands in sleeves, playful tilt, autumn park, falling maple leaves, golden hour lighting, cinematic shot

---

Negative Prompt Templates

Preset A: Standard Clean Anima DiT Negative (Recommended Default)

worst quality, low quality, score_1, score_2, score_3, blurry, jpeg artifacts, chromatic aberration, multiple girls, 2girls, group, crowd, bad hands, missing fingers, extra digits, cropped, frame, watermark, signature

Preset B: Anti-Wardrobe Bleed Negative (For Radical Outfit Changes)

Use when changing into swimsuits, armor, or fantasy gowns to suppress modern casual clothing:

worst quality, low quality, score_1, score_2, score_3, t-shirt, hoodie, sweater, jacket, jeans, sneakers, casual wear, modern clothes, extra limbs, bad anatomy, mutated hands, multiple views, split screen, collage

Preset C: Pure Face & Anatomy Protector (Close-ups & Portraits)

worst quality, low quality, score_1, score_2, score_3, distorted eyes, asymmetrical eyes, lazy eye, crossed eyes, bad mouth, extra teeth, disfigured face, poor skin texture, plastic skin, oversaturated, harsh shadows, multiple people

---

Technical Dataset & Architecture Notes

• Dataset Pool: 55 curated frames (30% Headshots, 40% Cowboy, 30% Full-Body).
• Network Parameters: Rank 32 / Alpha 16 (0.5 scaling ratio).
• Trained Steps: 1,650 steps at Batch Size 1 on an A100 Tensor Core GPU.