❄️ Odette (Genshin Impact) — Anima

Odette, the swan-ballerina, trained for Anima with full text-encoder training — so her design holds together under complex prompts, unusual lighting and poses that never appeared in the training data.

Particular attention went to the details most versions lose:

• the gold swan emblem on the corset

• the asymmetric gloves (one long blue glove, one bare arm)

• the ribbon-laced pointe heels

• the winged back ornament — correct from behind, not guessed

Trained at 1024px on curated in-game reference captures plus community art.

⚙️ Usage

Trigger: odette \(genshin impact\)

Strength: 1.0

Base: Anima Base v1.0 / waiANIMA

Steps: 25–30

CFG: 4–5

Sampler: er_sde / simple

The trigger is plain natural language, not an invented token. Write it exactly as above - escape the parentheses if your UI needs it.

No extra tags required. Hair colour, eye colour and her outfit all hold on their own. You don't need to describe her to get her, but might help to keep consistency.

💡 Prompt tips

Give it framing. Anima is somewhat NSFW-capable base and will drift toward suggestive angles on a bare 1girl, solo - this is the checkpoint's behaviour, not this LoRA's. Tell it what shot you want:

full body, standing, looking at viewer, simple background

Suggested negative:

nsfw, nude, ass focus, from below, panties, large breasts, worst quality, low quality, blurry, watermark, signature, text

For outdoor and scenery shots, add anime style, illustration to the positive and avoid stacking cinematic lighting with scenery - that combination pushes the base model toward a painterly game-cinematic look rather than clean anime.

Things that respond well: ballet pose, dancing, one leg raised, from behind, looking back, dynamic pose, snow, ice, winter.

🔧 Troubleshooting

Output looks like a game screenshot. Add anime style, illustration, 2d to the positive and drop cinematic lighting. Prefer day over night for outdoor scenes.

Composition is more suggestive than I asked for. Use the negative prompt above and add explicit framing tags. This is the base checkpoint, not the LoRA - it happens at LoRA strength 0 too.

Detail looks smeared. Drop to 0.85 strength, or raise steps to 30.

📝 Version notes

v1 - The model was trained on character reference captures, so outfit accuracy should be consistent.

Every design element is tagged separately in training, so pose, framing and lighting stay fully steerable without losing her identity.

Feedback and bug reports welcome - they genuinely improve the next version.