Preference Diffusion: Optimizing diffusion models with minimal human feedback
Diffusion-based image generation models (Stable Diffusion) are good at generating high-quality images but the user has little to no control over the output. To solve this, we employ Direct Preference Optimization (DPO), to align the prompt to the generated output images. We augment this method with real-time minimal human feedback in the form of their output preference, hence narrowing the image space before applying DPO using Gaussian Mixture Modeling.
https://drive.google.com/file/d/1qt0nLLlfGLfcmPBHBcgGhF4dSBFDt3kK/view?usp=sharing