This paper addresses the issue of dataset bias in diffusion models, which can significantly impact the quality and fairness of generated samples. The authors propose a method called Time-dependent Importance reWeighting (TIW) to mitigate this bias. TIW uses a time-dependent density ratio to estimate the bias between the biased and unbiased distributions, which is then used for both reweighting and score correction. This approach improves the accuracy of density ratio estimation and reduces error propagation in generative learning. The proposed method is theoretically connected to traditional score-matching objectives and is shown to converge to an unbiased distribution. Experimental results on various datasets, including CIFAR-10, CIFAR-100, FFHQ, and CelebA, demonstrate the effectiveness of TIW in reducing bias and improving sample quality compared to baseline methods. The code for the proposed method is available at <https://github.com/alsdudrla10/TIW-DSM>.This paper addresses the issue of dataset bias in diffusion models, which can significantly impact the quality and fairness of generated samples. The authors propose a method called Time-dependent Importance reWeighting (TIW) to mitigate this bias. TIW uses a time-dependent density ratio to estimate the bias between the biased and unbiased distributions, which is then used for both reweighting and score correction. This approach improves the accuracy of density ratio estimation and reduces error propagation in generative learning. The proposed method is theoretically connected to traditional score-matching objectives and is shown to converge to an unbiased distribution. Experimental results on various datasets, including CIFAR-10, CIFAR-100, FFHQ, and CelebA, demonstrate the effectiveness of TIW in reducing bias and improving sample quality compared to baseline methods. The code for the proposed method is available at <https://github.com/alsdudrla10/TIW-DSM>.