xlsd32-alpha1
| Property | Value |
|---|---|
| Author | opendiffusionai |
| Model Type | Stable Diffusion 1.5 with SDXL VAE |
| Training Precision | FP32 |
| Model URL | Hugging Face |
What is xlsd32-alpha1?
xlsd32-alpha1 is an experimental work-in-progress model that combines Stable Diffusion 1.5 base with the SDXL VAE, specifically retrained to achieve compatibility and performance. Currently at epoch 10, the model has reportedly reached parity with existing SD base models for human output generation.
Implementation Details
The model was trained using full fp32 precision on a single NVIDIA 4090 GPU. The training dataset comprises multiple LAION2B variants, including both 1120px and 1024px/1536px square images, annotated with both moondream and wd14 captions.
- Training datasets include LAION2B combinations at 1120px, 1024px, and 1536px resolutions
- Utilizes both moondream and wd14 caption systems
- Trained with full FP32 precision for maximum accuracy
- Incorporates SDXL VAE architecture with SD1.5 base model
Core Capabilities
- Compatible with standard SD1.5 workflows and pipelines
- Optimized for human image generation
- Supports high-resolution image generation
- Combines benefits of SD1.5 and SDXL architectures
Frequently Asked Questions
Q: What makes this model unique?
This model uniquely combines SD1.5's base architecture with SDXL's VAE, creating a hybrid approach that maintains compatibility with SD1.5 while potentially offering improved image quality through the SDXL VAE.
Q: What are the recommended use cases?
The model is designed to be used like any other SD1.5 model, with particular strength in human image generation. It's currently in alpha stage, making it suitable for experimental use and testing rather than production environments.
