EMoE: Training-Free Expert Disagreement for Uncertainty-Aware Text-to-Image Diffusion
Researchers have developed a new method called EMoE to estimate uncertainty in text-to-image diffusion models without requiring additional training. EMoE leverages the disagreement between different 'expert' pathways within existing Mixture-of-Experts (MoE) diffusion models. By measuring the variance in latent representations after the first denoising step, EMoE can predict the likelihood of a poorly aligned image generation, offering a practical tool for analyzing prompt risk and model biases. AI
IMPACT Provides a training-free method to assess prompt risk and model biases in diffusion models.