FBP-diffusion: diffusion model combining MobileViT and dynamic loss correction for facial beauty prediction

Facial beauty prediction (FBP) is a frontier topic at the intersection of artificial intelligence and computational aesthetics, aiming to enable computers to autonomously predict or assess facial beauty. Currently, while FBP methods have achieved good results on well-processed datasets, they typically exhibit reduced prediction performance on datasets with more unavoidable noisy labels. Diffusion models (DMs) can denoise and reconstruct label encodings, capturing uncertainty in the prediction process through the randomness of their outputs. Therefore, we propose FBP-Diffusion, an improved diffusion model that integrates MobileViT and dynamic loss correction (DLC). Specifically, MobileViT, effective at modeling both detailed and global information, is employed as a conditional information encoder to produce preliminary predictions, which are then fed into the reverse process to guide label generation. DLC is introduced to enhance the model’s denoising capability and robustness, in which the cross-entropy loss is increased by the prediction probabilities of FBP-Diffusion obtained after the reverse process and probability transfer, and then dynamically integrated into the noise estimation loss. Experimental results on four representative facial beauty databases demonstrate that FBP-Diffusion outperforms both conventional DMs and FBP methods, particularly noting a 5.17% accuracy improvement on relatively noisy datasets over state-of-the-art FBP methods.