kyoStyle LoRA - Anima

A dedicated style adaptation designed specifically for the Anima architecture. This model is fine-tuned to capture the intricate, and vibrant visual DNA of Kyoto Animation.

Trigger Tags

To activate the style, include these tags at the beginning of your prompt:

anime, kyoStyle

⚠️ IMPORTANT: Character Disclaimer

NO SPECIFIC CHARACTERS WERE TRAINED. This LoRA contains zero character personality or identity data.

Every single character you see in the showcase gallery (Rei, Asuka, Shinji, etc.) is already baked into the Anima base model.

This LoRA does NOT teach the model who they are; it merely reshapes their geometry, lines, and lighting into the KyoStyle aesthetic. If the base model knows the character via Natural Language or tags, this LoRA will seamlessly skin them.

Dataset

High-Quality Large Screencaps: Trained on a meticulously curated mix of high-resolution screencaps and synthetic data.

Subpixel Mastery: The dataset was tailored to improve fine details like eye irises, wet/teary eye reflections, and continuous, clean hair line-art that doesn't break or artifact.

Captioning Workflow: All captions and datasets were managed and refined using my custom offline tool: TextManager on GitHub (made dataset curation for this architecture a breeze).

Training Specifications.

If you want to replicate or understand the behavior of this LoRA, here is the exact technical pipeline used:

Trainer: Anima-Standalone-Trainer, base model

Network Rank (Dim/Alpha): 32 / 32
Stage 1 (Base Training): 10,000 steps at 768px to capture core composition, color theory, and macro-features of the style.
Stage 2 (High-Res Finetune): 1,000 extra steps at 1024px with an ultra-low Learning Rate and frozen Text Encoder. This progressive rescaling pushed the model to utilize the extra pixels purely for micro-details, line smoothing, and advanced eye/shading rendering.

Permissions & License

✅ Finetuning allowed.

✅ Merges allowed.

✅ Commercial use allowed for any purposes (except selling this model).