RAVE: Residual Vector Embedding for Efficient CLIP-Guided Backlit Image Enhancement

This work proposes two novel methods, CLIP-LIT-Latent and RAVE, that efficiently utilize CLIP guidance for backlit image enhancement. CLIP-LIT-Latent trains vectors directly in the CLIP latent space, while RAVE computes a residual vector in the CLIP embedding space to guide the enhancement model, leading to faster training and higher quality results compared to the original CLIP-LIT approach.