Diffusion models (DMs) have achieved significant success in generating
imaginative images given textual descriptions. However, they are likely to fall
short when it comes to real-life scenarios with intricate details.The
low-quality, unrealistic human faces in text-to-image generation are one of the
most prominent issues, hindering the wide application of DMs in practice.
Targeting addressing such an issue, we first assess the face quality of
generations from popular pre-trained DMs with the aid of human annotators and
then evaluate the alignment between existing metrics such as ImageReward, Human
Preference Score, Aesthetic Score Predictor, and Face Quality Assessment, with
human judgments. Observing that existing metrics can be unsatisfactory for
quantifying face quality, we develop a novel metric named Face Score (FS) by
fine-tuning ImageReward on a dataset of (good, bad) face pairs cheaply crafted
by an inpainting pipeline of DMs. Extensive studies reveal that FS enjoys a
superior alignment with humans. On the other hand, FS opens up the door for
refining DMs for better face generation. To achieve this, we incorporate a
guidance loss on the denoising trajectories of the aforementioned face pairs
for fine-tuning pre-trained DMs such as Stable Diffusion V1.5 and Realistic
Vision V5.1. Intuitively, such a loss pushes the trajectory of bad faces toward
that of good ones. Comprehensive experiments verify the efficacy of our
approach for improving face quality while preserving general capability.

通过评估现有度量标准并开发一种新的面部评分指标（Face Score），本研究提出了一种改进扩散模型以提高生成图像质量的方法。引入指导损失函数用于微调预训练的扩散模型，实验证实了该方法在提高面部质量的同时保持了其一般能力。