| Property | Value |
|---|---|
| License | CreativeML OpenRAIL-M |
| Author | alea31415 |
| Training Dataset Size | 27K images |
| Training Cost | ~$15 (RunPod) |
What is bocchi-the-rock-character?
This is a specialized AI model designed to generate character images from the anime "Bocchi the Rock!" The model's unique feature is its ability to create multi-character scenes, particularly focusing on the Kessoku Band members and other characters from the series. It supports both single-character generation and complex multi-character compositions, with the added capability of inpainting for refinement.
Implementation Details
The model was trained using EveryDream1 trainer for 50,000 steps with a batch size of 4, learning rate of 1e-6, and 512 resolution. The training dataset comprises 7,024 anime screenshots, 1,630 fan arts, and 18,519 customized regularization images, weighted at 0.3, 0.25, and 0.45 respectively.
- Training Infrastructure: RunPod with NVIDIA 3090 GPU
- Conditional dropping rate: 10%
- Character support: 12 distinct characters
- Custom weighting scheme for balanced concept learning
Core Capabilities
- Multi-character scene generation
- Individual character rendering with appearance customization
- Inpainting support for scene refinement
- Balanced representation of different art styles
Frequently Asked Questions
Q: What makes this model unique?
The model's ability to generate multi-character scenes and its specialized training on Bocchi the Rock characters sets it apart. The weighted training approach ensures balanced representation across different image types.
Q: What are the recommended use cases?
The model is ideal for generating fan art, character compositions, and scene recreation from Bocchi the Rock. It's particularly effective when combined with inpainting for refined results.