This is a high-fidelity SDXL LoRA trained to generate Zero Two (ゼロツー) from the anime Darling in the Franxx. It was trained on 95 curated images with Danbooru-style captions, covering a wide range of poses, expressions, outfits, and scenarios.

The model captures Zero Two's signature features — long pink hair, cyan/turquoise eyes, and red horns — across various settings including her red military uniform, white pilot suit, casual wear, and swimwear.

What is this model good at:

• Generating recognizable, consistent depictions of Zero Two across diverse poses and outfits

• Maintaining her distinctive character features (pink hair, horns, cyan eyes)

• Working well with Danbooru-style tag prompts for fine-grained control over pose, expression, and clothing

• Beach, outdoor, action, and portrait compositions

What is this model not good at:

• Generating other characters — this is a single-character LoRA

• Highly photorealistic styles (it was trained on anime artwork)

• Multi-character scenes (trained primarily on solo images)

• Non-anime art styles

Trigger Words & Usage

To use this model, include the following in your prompt: zerotwo, darling in the franxx, 1girl

• zerotwo is the main trigger word for the character.

• darling in the franxx reinforces the source material style.

• 1girl is included for stability and compatibility with the Animagine XL base model.

• solo can also be added for single character focus.

Recommended Settings:

• Base Model: Animagine XL V3.1 (https://huggingface.co/cagliostrolab/animagine-xl-3.1)

• LoRA Weight: 0.8 (can be adjusted between 0.7–0.9 for desired strength)

• CFG Scale: 6.5

• Steps: 30

• Sampler: Euler a (Euler Ancestral Discrete Scheduler) or DPM++ 2M Karras

• Resolution: 1024x1024

Prompting Advice:

The model was trained using Danbooru-style tags. For best results, use descriptive tags for features, clothing, expressions, and poses (e.g., red military uniform, white bodysuit, looking at viewer, smile, full body, outdoors, forest).

For enhanced quality, use tags like masterpiece, best quality.

Using negative prompts is highly recommended. A good starting point is: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, jpeg artifacts, signature, watermark, username, blurry, adult, photorealistic, simple eyes, dot eyes, nsfw

Training Details:

• Network Dimensions: 32 / Alpha: 16

• Epochs: 8

• Learning Rate: 5e-5 (UNet only, text encoder frozen)

• Optimizer: AdamW8bit with cosine LR scheduler

• Dataset: 95 images × 10 repeats

• Min SNR Gamma: 5.0

• Mixed Precision: bf16