中国科学院自动化研究所机构知识库

As an extension of diffusion models, consistency models reduce the necessary sampling steps to a single iteration when synthesizing an image sample, thereby significantly enhances the efficiency of image generation. In addition, they also allow multi-step generation, providing flexibility for trade-offs between sample quality and computational efficiency. Despite these advantages, traditional consistency models are troubled by loss of high-frequency image details. This issue is attributed to the inherent regression-to-mean property of the L2 training loss, impeding overall improvement of model performance on the image generation task. In this paper, we propose a novel consistency adversarial model to address the loss of high-frequency image details through adversarial generation. In particular, we train a consistency model in an adversarial manner by treating it as an image generator. Then, an additional image discriminator is introduced and optimized along with the consistency model in an adversarial manner. The target of the image discriminator is to punish the image generator when it synthesizes images lacking high-frequency image details. In this way, image samples with high-frequency image details can be obtained and the performance of consistency models can be improved. Extensive experiments demonstrate the effectiveness of our proposed method. Our CAM outperforms the traditional consistency model on two challenging benchmarks: ImageNet and LSUN.