| Property | Value |
|---|---|
| License | CreativeML OpenRAIL-M |
| Type | Text-to-Image |
| Authors | Xerxemi & Xynon |
| Format | Safetensors |
What is SD-Silicon?
SD-Silicon represents a breakthrough series of general-purpose models developed using the experimental automerger, autoMBW. Created through a collaboration between Xerxemi and Xynon, these models feature built-in WD1.3 VAE capabilities while maintaining compatibility with external VAE implementations.
Implementation Details
The series includes multiple variants, with Silicon28 (extestg4) being the first to achieve quality comparable to manual merge block weight merges. Silicon29 (extesto4) builds upon this foundation with an expanded merge list and improved stability. The models also feature specialized versions like Silicon28-negzero (negatively finetuned) and Silicon29-dark (merged with noise offset for darker outputs).
- Integrated WD1.3 VAE architecture
- Automated merge block weight technology
- Multiple specialized variants for different use cases
- Optimized performance with recommended settings
Core Capabilities
- High-quality text-to-image generation
- Versatile general-purpose implementation
- Specialized dark and negative variants
- Optimal performance with DPM++ 2M sampler
Frequently Asked Questions
Q: What makes this model unique?
SD-Silicon stands out for its use of autoMBW technology, achieving quality levels that match or exceed manual merge block weight approaches. It's the first successful implementation of automated model merging at this scale.
Q: What are the recommended use cases?
The model excels in general-purpose image generation with recommended settings including DPM++ 2M sampler, 42+42 steps, Latent upscaler, 0.5-0.6 denoising, and CFG 13. Different variants cater to specific needs like darker outputs or specialized negative generations.