New
Stability Generative Models
- Release
SD-XL 0.9-baseandSD-XL 0.9-refinermodels - Release new
sgmcodebase - Remove extensive subclassing
- Handle all types of conditioning inputs (vectors, sequences and spatial conditionings, and all combinations thereof) in a single class
GeneralConditioner - Separate guiders and samplers independent of the model
- Adopt the "denoiser framework" for both training and inference
- Discrete times models are special case of continuous time models
- Make independent following features: weighting of the diffusion loss function, preconditioning of the network, and sampling of noise levels during training
- Clean up Autoencoding models