This paper emphasizes the importance of color in dataset condensation and proposes DC3 (Dataset Condensation with Color Compensation), a novel method that overcomes the shortcomings of existing methods. We point out that existing image-level selection methods (Coreset Selection and Dataset Quantization) lead to inefficient reduction, and pixel-level optimization methods (Dataset Distillation) cause semantic distortion due to excessive parameters. DC3 utilizes a latent diffusion model to enhance the color diversity of existing images instead of generating new images after a compensated selection strategy. We demonstrate that DC3 is the first study to fine-tune a pre-trained diffusion model on a reduced dataset, outperforming existing state-of-the-art (SOTA) methods in various benchmarks and generalizing well. Our FID results demonstrate that network training on a high-quality dataset is possible without model collapse or performance degradation.
