<p>In the field of imbalanced image generation, long-tailed diffusion models have achieved outstanding results. However, there has been little research on the exposure bias problem that exists in long-tailed diffusion models. In this paper, we conduct a systematic study of the exposure bias in long-tailed diffusion models and find that this bias arises from the inconsistency in network inputs between the training and sampling phases. To mitigate this bias, we propose a method that requires no additional training and improves performance from two aspects: time-step selection and mean estimation. Regarding time-step selection, our study reveals that during sampling, for a given time-step <i>t</i> and its corresponding <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$\hat{x}_t$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mover accent="true"> <mi>x</mi> <mo stretchy="false">^</mo> </mover> <mi>t</mi> </msub> </math></EquationSource> </InlineEquation>, there may exist an alternative time-step <InlineEquation ID="IEq2"> <EquationSource Format="TEX">$t_k$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>t</mi> <mi>k</mi> </msub> </math></EquationSource> </InlineEquation> that exhibits a stronger coupling with <InlineEquation ID="IEq3"> <EquationSource Format="TEX">$\hat{x}_t$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mover accent="true"> <mi>x</mi> <mo stretchy="false">^</mo> </mover> <mi>t</mi> </msub> </math></EquationSource> </InlineEquation>. Based on this observation, we adjust the sampling time-steps accordingly. Regarding mean estimation, we observe that in the sampling process of long-tailed diffusion models, there is gradient information between the estimated values of the original sample at two consecutive time-steps. By leveraging this gradient information, we extrapolate the two estimates to compute a more accurate overall estimate. Our method can be applied to various long-tailed diffusion models (e.g., CBDM, T2H) and is compatible with multiple samplers, including DDPM, DDIM, and other higher-order samplers. For instance, on the CIFAR100LT long-tailed dataset, our method achieves an impressive FID of 5.73 when using the DEIS sampler. The link to our code is: <a href="https://github.com/king-w-king/CBDM_TS">https://github.com/king-w-king/CBDM_TS</a>.</p>

Mitigating exposure bias in long-tailed diffusion models through improved time-steps and mean estimation

期刊论文
发表日期：

2026年5月8日

查看全文

Zhuoyu Zhou,
Qiangkui Leng,
Keyi Song,
Guansheng Yuan

摘要

In the field of imbalanced image generation, long-tailed diffusion models have achieved outstanding results. However, there has been little research on the exposure bias problem that exists in long-tailed diffusion models. In this paper, we conduct a systematic study of the exposure bias in long-tailed diffusion models and find that this bias arises from the inconsistency in network inputs between the training and sampling phases. To mitigate this bias, we propose a method that requires no additional training and improves performance from two aspects: time-step selection and mean estimation. Regarding time-step selection, our study reveals that during sampling, for a given time-step t and its corresponding $\hat{x}_t$ ${\hat{x}}_{t}$ , there may exist an alternative time-step $t_k$ $t_{k}$ that exhibits a stronger coupling with $\hat{x}_t$ ${\hat{x}}_{t}$ . Based on this observation, we adjust the sampling time-steps accordingly. Regarding mean estimation, we observe that in the sampling process of long-tailed diffusion models, there is gradient information between the estimated values of the original sample at two consecutive time-steps. By leveraging this gradient information, we extrapolate the two estimates to compute a more accurate overall estimate. Our method can be applied to various long-tailed diffusion models (e.g., CBDM, T2H) and is compatible with multiple samplers, including DDPM, DDIM, and other higher-order samplers. For instance, on the CIFAR100LT long-tailed dataset, our method achieves an impressive FID of 5.73 when using the DEIS sampler. The link to our code is: https://github.com/king-w-king/CBDM_TS.