DMD2MOD
| Property | Value |
|---|---|
| Author | YOB-AI |
| Original Model | DMD2 by tianweiy |
| Repository | Hugging Face |
What is DMD2MOD?
DMD2MOD is an innovative adaptation of the original DMD2 trajectory consistency model, specifically designed to enhance compatibility with various samplers and higher CFG (Classifier Free Guidance) settings. This model represents a significant advancement in addressing the limitations of the base DMD2 model, particularly focusing on problematic scales that emerge when using high CFG values.
Implementation Details
The model introduces a modified approach to trajectory consistency, featuring LORA-64-dmd2.safetensors as its core component. This implementation includes a base extraction of DMD (step 0) with carefully calibrated weight influence reduction, ensuring better performance across different sampling methods.
- Enhanced compatibility with LCM and Kohaku samplers
- Optimized for 1-2 CFG settings with potential for higher values
- Reconstructed scale handling for improved stability
- Tested compatibility with NoobAI v.1.0
Core Capabilities
- Extended sampler compatibility beyond original DMD2 limitations
- Improved handling of high CFG scenarios
- Reduced weight influence for better stability
- Maintained trajectory consistency while expanding use cases
Frequently Asked Questions
Q: What makes this model unique?
DMD2MOD stands out for its ability to work with multiple samplers while maintaining stability at higher CFG values, addressing key limitations of the original DMD2 model. The carefully implemented scale reconstruction and weight adjustments make it particularly valuable for advanced AI applications.
Q: What are the recommended use cases?
This model is particularly suited for applications requiring trajectory consistency with various samplers, especially when working with LCM and Kohaku samplers. It's ideal for testing and experimental scenarios where standard DMD2 capabilities need to be extended beyond their usual limitations.
