ConsistencyDecoder is a Python package from OpenAI that introduces an improved decoding method for variational autoencoders (VAEs) used in Stable Diffusion pipelines. Instead of relying solely on the standard GAN or VAE decoder, this approach leverages a Consistency Distilled Diff VAE, designed to produce higher-quality and more stable outputs from encoded latents. The project provides a simple API for encoding with a Stable Diffusion VAE and decoding using the new consistency model, allowing for side-by-side comparisons with traditional decoders. It demonstrates how consistency models can enhance visual fidelity while maintaining efficiency, reducing artifacts common in GAN-decoded outputs. The repository includes installation instructions, usage examples, and visual comparisons to highlight improvements. Though compact, it provides an accessible entry point for experimenting with advanced decoding strategies in diffusion-based generative models.
Features
- Consistency Distilled Diff VAE for improved decoding in Stable Diffusion
- Compatible with Hugging Face Diffusers StableDiffusionPipeline
- Provides side-by-side comparisons of GAN, VAE, and consistency decoders
- Simple API for encoding latents and decoding with the new method
- Example images included for testing and reproducibility
- Lightweight Python package installable directly from GitHub