Kohaku-XL-Delta
| Property | Value |
|---|---|
| License | FAIR AI Public License 1.0 |
| Architecture | StableDiffusionXL Pipeline |
| Training Dataset Size | 3.6M images |
| Research Paper | LyCORIS Fine-tuning Paper |
What is Kohaku-XL-Delta?
Kohaku-XL-Delta is the fourth major iteration in the Kohaku XL series, representing a significant advancement in anime-style image generation. Trained on a massive dataset of 3.6 million images using LyCORIS fine-tuning, this model was developed on consumer-grade hardware while maintaining professional-grade output quality.
Implementation Details
The model utilizes the LoKr algorithm with full matrix triggering and employs factors of 2-8 for different modules. Training was conducted on dual RTX 3090s over approximately 17-18 days, processing 28,638 steps with an equivalent batch size of 128.
- Trained using Lion8bit optimizer with carefully tuned learning rates
- Implements mixed precision FP16 training
- Supports resolutions from 256x256 up to 4096x4096
- Features advanced tag system with quality, rating, and date categorization
Core Capabilities
- Sophisticated tag handling with support for all Danbooru tags (1000+ popularity)
- Quality-based image generation with 7 distinct quality levels
- Specialized artist style blending capabilities
- Support for various aspect ratios and high-resolution outputs
Frequently Asked Questions
Q: What makes this model unique?
The model's distinctive feature is its comprehensive training on a carefully curated dataset, combined with advanced LyCORIS fine-tuning techniques. It offers superior anime-style image generation while maintaining the characteristic look of AI-generated art.
Q: What are the recommended use cases?
The model excels at generating anime-style images with specific quality levels and artistic styles. It's particularly effective when blending multiple artist tags rather than replicating specific artists' styles.
