The paper introduces a novel approach called DiffClass for exemplar-free class incremental learning (CIL) to address the challenges of catastrophic forgetting and domain gap issues. The method uses multi-distribution matching (MDM) diffusion models to align the quality of synthetic data and bridge domain gaps among all training domains. Additionally, selective synthetic image augmentation (SSIA) is integrated to expand the training data distribution, enhancing the model's plasticity and stability. The approach reformulates exemplar-free CIL as a multi-domain adaptation (MDA) problem to implicitly manage domain gaps during incremental training. Extensive experiments on benchmark datasets demonstrate that DiffClass outperforms existing exemplar-free CIL methods with significant improvements, achieving state-of-the-art performance. The project page is available at https://cr8br0ze.github.io/DiffClass.The paper introduces a novel approach called DiffClass for exemplar-free class incremental learning (CIL) to address the challenges of catastrophic forgetting and domain gap issues. The method uses multi-distribution matching (MDM) diffusion models to align the quality of synthetic data and bridge domain gaps among all training domains. Additionally, selective synthetic image augmentation (SSIA) is integrated to expand the training data distribution, enhancing the model's plasticity and stability. The approach reformulates exemplar-free CIL as a multi-domain adaptation (MDA) problem to implicitly manage domain gaps during incremental training. Extensive experiments on benchmark datasets demonstrate that DiffClass outperforms existing exemplar-free CIL methods with significant improvements, achieving state-of-the-art performance. The project page is available at https://cr8br0ze.github.io/DiffClass.