Conditioning diffusion models via attributes and semantic masks for face generation

Deep generative models have shown impressive results in generating realistic images of faces. GANs managed to generate high-quality, high-fidelity images when conditioned on semantic masks, but they still lack the ability to diversify their output. Diffusion models partially solve this problem and are able to generate diverse samples given the same condition. This paper introduces a novel strategy for enhancing diffusion models through multi-conditioning, harnessing cross-attention mechanisms to utilize multiple feature sets, ultimately enabling the generation of high-quality and controllable images. The proposed method extends previous approaches by introducing conditioning on both attributes and semantic masks, ensuring finer control over the generated face images. In order to improve the training time and the generation quality, the impact of applying perceptual-focused loss weighting into the latent space instead of the pixel space is also investigated. The proposed solution has been evaluated on the CelebA-HQ dataset, and it can generate realistic and diverse samples while allowing for fine-grained control over multiple attributes and semantic regions. Experiments on the DeepFashion dataset have also been performed in order to analyze the capability of the proposed model to generalize to different domains. In addition, an ablation study has been conducted to evaluate the impact of different conditioning strategies on the quality and diversity of the generated images.