Suppress-to-Impress: Attenuating Distractions in Attention Maps for Diffusion Model Inference
2023-06-28 ~ now
Abstract I am focusing on harnessing the power of diffusion models for conditional image generation. Essentially, I am aiming to guide the cross-attention to focus only on what’s vital by suppressing the less relevant parts in the cross attention map. To break it down, conditional image generation is all about...