**DiffuseMix: Label-Preserving Data Augmentation with Diffusion Models** **Abstract:** DiffuseMix is a novel data augmentation technique that leverages diffusion models to generate diverse and label-preserving augmented images. Unlike traditional image-mixing methods that can omit important portions of the input images and introduce label ambiguities, DiffuseMix employs conditional prompts to generate hybrid images by concatenating a portion of the original image with its generated counterpart. A random fractal image is then blended into the hybrid image to enhance structural diversity and improve adversarial robustness. Empirical results on seven datasets demonstrate that DiffuseMix outperforms existing state-of-the-art methods in tasks such as general classification, fine-grained classification, fine-tuning, data scarcity, and adversarial robustness. **Introduction:** Image-mixing-based augmentation techniques have been widely used to improve the generalization of deep neural networks. However, these methods can omit important image regions and introduce label ambiguities. DiffuseMix addresses these limitations by generating diverse images using tailored conditional prompts and blending them with fractal images to enhance structural complexity and avoid overfitting. **Method:** DiffuseMix consists of three main steps: generation, concatenation, and fractal blending. Conditional prompts are used to generate a generative counterpart of the input image. A portion of the original image is concatenated with the generated counterpart using a binary mask to create a hybrid image. Finally, a random fractal image is blended into the hybrid image to form the final augmented image. **Experiments and Results:** DiffuseMix is evaluated on various datasets, including ImageNet, CIFAR-100, Tiny-ImageNet-200, Oxford-102 Flower, Stanford Cars, Aircraft, and Caltech-UCSD Birds-200-2011. Results show that DiffuseMix achieves superior performance compared to existing methods in general classification, fine-grained classification, adversarial robustness, transfer learning, and data scarcity tasks. **Conclusion:** DiffuseMix effectively increases the diversity of training data while preserving the original semantics of the input images. It demonstrates consistent performance gains across multiple tasks and benchmark datasets, outperforming existing state-of-the-art image augmentation methods.**DiffuseMix: Label-Preserving Data Augmentation with Diffusion Models** **Abstract:** DiffuseMix is a novel data augmentation technique that leverages diffusion models to generate diverse and label-preserving augmented images. Unlike traditional image-mixing methods that can omit important portions of the input images and introduce label ambiguities, DiffuseMix employs conditional prompts to generate hybrid images by concatenating a portion of the original image with its generated counterpart. A random fractal image is then blended into the hybrid image to enhance structural diversity and improve adversarial robustness. Empirical results on seven datasets demonstrate that DiffuseMix outperforms existing state-of-the-art methods in tasks such as general classification, fine-grained classification, fine-tuning, data scarcity, and adversarial robustness. **Introduction:** Image-mixing-based augmentation techniques have been widely used to improve the generalization of deep neural networks. However, these methods can omit important image regions and introduce label ambiguities. DiffuseMix addresses these limitations by generating diverse images using tailored conditional prompts and blending them with fractal images to enhance structural complexity and avoid overfitting. **Method:** DiffuseMix consists of three main steps: generation, concatenation, and fractal blending. Conditional prompts are used to generate a generative counterpart of the input image. A portion of the original image is concatenated with the generated counterpart using a binary mask to create a hybrid image. Finally, a random fractal image is blended into the hybrid image to form the final augmented image. **Experiments and Results:** DiffuseMix is evaluated on various datasets, including ImageNet, CIFAR-100, Tiny-ImageNet-200, Oxford-102 Flower, Stanford Cars, Aircraft, and Caltech-UCSD Birds-200-2011. Results show that DiffuseMix achieves superior performance compared to existing methods in general classification, fine-grained classification, adversarial robustness, transfer learning, and data scarcity tasks. **Conclusion:** DiffuseMix effectively increases the diversity of training data while preserving the original semantics of the input images. It demonstrates consistent performance gains across multiple tasks and benchmark datasets, outperforming existing state-of-the-art image augmentation methods.