MF-Base

Maintained By
MyneFactory

MF-Base: Myne Factory Base Model

PropertyValue
LicenseCreativeML OpenRAIL-M
Authors金Goldkoron, tsmkirby, Juusoz
Training Data~23,000 high-scored samples

What is MF-Base?

MF-Base is a sophisticated text-to-image model designed as the foundation of Myne Factory's model ecosystem. Built using approximately 18,000 high-scored samples from Yande.re and 5,000 from Konachan, this model represents a significant advancement in anime-style image generation capabilities.

Implementation Details

The model employs a unique training approach using three iterations of WD1.4 tagger for comprehensive object identification within the training data. Training was conducted on an RTX 4090 with a batch size of 4, utilizing DDIM sample scheduler and DDPM noise scheduler with mix precision. The text encoder underwent training for 50% of the duration, with alternating freeze-unfreeze cycles every 10 epochs.

  • Advanced caption generation system with multiple tagger iterations
  • Optimized using Adam optimizer with manual learning rate control
  • Specialized text encoder training methodology
  • Custom implementation of DDIM and DDPM schedulers

Core Capabilities

  • High-quality anime-style image generation
  • Superior handling of caustics and depth effects
  • Efficient processing of both detailed and shorter prompts
  • Versatile resolution support (1024x576 to 768x768)

Frequently Asked Questions

Q: What makes this model unique?

MF-Base stands out for its specialized training on highly-curated anime content and its advanced handling of visual effects like caustics and depth of field. The model's ability to process both complex and simple prompts makes it versatile for various use cases.

Q: What are the recommended use cases?

The model excels at generating anime-style artwork, particularly with booru-style tags. It's optimized for various aspect ratios and performs exceptionally well with both character-focused and landscape compositions. The recommended settings include using DPM++ SDE Karras sampler with 30+ steps and CFG scales between 9-12.

🍰 Interesting in building your own agents?
PromptLayer provides Huggingface integration tools to manage and monitor prompts with your whole team. Get started here.