Multi-Step Denoising Scheduled Sampling: Towards Alleviating Exposure Bias for Diffusion Models

Zhiyao Ren, Yibing Zhan, Liang Ding, Gaoang Wang, Chaoyue Wang, Zhongyi Fan, Dacheng TaoAAAI 2024 · OpenReview

Abstract

Denoising Diffusion Probabilistic Models (DDPMs) have achieved significant success in generation tasks. Nevertheless, the exposure bias issue—the natural discrepancy between training (where inputs are ground-truth derived) and inference (where inputs are iteratively generated by the model)—harms performance. We propose Multi-Step Denoising Scheduled Sampling (MDSS) to alleviate this bias. MDSS comprehensively considers the influence of prediction errors on both the model's noise output and the resulting input signal for the next step. By efficiently modeling multi-step error accumulation through a mathematical formulation, MDSS enables the model to learn self-correction. Experimental results show that MDSS effectively mitigates exposure bias in DDPM, DDIM, and DPM-solver, achieving a state-of-the-art FID score of 3.86 on CIFAR-10 with 100 DDIM steps.

Introduction: The Exposure Bias Problem

Denoising Diffusion Probabilistic Models (DDPMs) generate images by starting with pure noise and iteratively removing it. However, there is a fundamental mismatch between how they are trained and how they are used, known as exposure bias.

The discrepancy between training and inference in DDPMs.

During Training, the model is shown a "clean" noisy image $x_t$ derived directly from a ground-truth image $x_0$. It learns to predict the noise added at that specific step. During Inference, the model starts from random noise and uses its own previous (imperfect) predictions to generate the input for the next step. Small errors in early steps compound over time, leading to a "drift" from the data distribution seen during training. While common in text generation, this issue has been largely overlooked in diffusion models until now.

Methodology: Multi-Step Denoising Scheduled Sampling (MDSS)

MDSS improves upon standard training by simulating realistic multi-step error accumulation, allowing the model to "see" and correct for its own mistakes during the training phase. The method introduces three key innovations:

  • Simulating Real Errors: Instead of adding random Gaussian noise, MDSS uses the model's actual one-step prediction to introduce realistic "model-based" errors into the training samples.
  • Multi-Step Accumulation: To simulate how errors compound without the massive computational cost of full iterative loops, the authors developed a mathematical formulation using the posterior distribution. This allows the model to efficiently calculate what a sample would look like after $k$ steps of imperfect inference.
  • Dual-Channel Denoising: The model is modified to predict two components: the noise in the current step (standard) and a "correction" signal to denoise the input signal itself. This helps the model handle both its current prediction error and errors inherited from previous steps.

Crucially, MDSS can be applied as a fine-tuning step on top of pre-trained models, making it a computationally efficient enhancement.

Experiments and Results

The researchers evaluated MDSS on standard benchmarks: CIFAR-10, ImageNet 64×64, and LSUN Tower 64×64. The results show that MDSS consistently outperforms previous state-of-the-art methods like Input Perturbation (IP) and standard Scheduled Sampling (SS).

Error Mitigation Analysis

The core strength of MDSS is its ability to maintain low Mean Squared Error (MSE) even when inputs are corrupted by accumulated errors.

MSE comparison across different input error scenarios: (a) no errors, (b) one-step errors, and (c) accumulated errors.

As shown in the figure above, MDSS ("Ours") consistently maintains significantly lower MSE across the sampling trajectory compared to baselines, particularly in the most challenging scenario (c) where errors have accumulated over multiple steps.

Quantitative Performance

  • Superior Image Quality: On CIFAR-10 using the DDIM sampler (100 steps), MDSS achieved a Frechet Inception Distance (FID) of 3.86, significantly better than the second-best score of 4.78.
  • Massive Gains in Low-Step Regimes: Exposure bias is most damaging when using fast samplers like DPM-Solver. With only 10 steps, MDSS improved the FID from 26.08 (standard) to 12.11, effectively doubling performance in ultra-fast generation scenarios.

Conclusion

MDSS provides a robust framework for alleviating exposure bias by modeling multi-step error accumulation. It is particularly effective for fast-sampling regimes (DDIM, DPM-Solver), making it a versatile "plug-and-play" enhancement for modern diffusion pipelines. Future work will explore adjustments for higher-order solvers and non-additive noise distributions.

Summary updated

Abstract

Denoising Diffusion Probabilistic Models (DDPMs) have achieved significant success in generation tasks. Nevertheless, the exposure bias issue, i.e., the natural discrepancy between the training (the output of each step is calculated individually by a given input) and inference (the output of each step is calculated based on the input iteratively obtained based on the model), harms the performance of DDPMs. To our knowledge, few works have tried to tackle this issue by modifying the training process for DDPMs, but they still perform unsatisfactorily due to 1) partially modeling the discrepancy and 2) ignoring the prediction error accumulation. To address the above issues, in this paper, we propose a multi-step denoising scheduled sampling (MDSS) strategy to alleviate the exposure bias for DDPMs. Analyzing the formulations of the training and inference of DDPMs, MDSS 1) comprehensively considers the discrepancy influence of prediction errors on the output of the model (the Gaussian noise) and the output of the step (the calculated input signal of the next step), and 2) efficiently models the prediction error accumulation by using multiple iterations of a mathematical formulation initialized from one-step prediction error obtained from the model. The experimental results, compared with previous works, demonstrate that our approach is more effective in mitigating exposure bias in DDPM, DDIM, and DPM-solver. In particular, MDSS achieves an FID score of 3.86 in 100 sample steps of DDIM on the CIFAR-10 dataset, whereas the second best obtains 4.78. The code will be available on GitHub.
BibTeX
@inproceedings{ren2024multi,
  author = {Zhiyao Ren and Yibing Zhan and Liang Ding and Gaoang Wang and Chaoyue Wang and Zhongyi Fan and Dacheng Tao},
  title = {{Multi-Step Denoising Scheduled Sampling: Towards Alleviating Exposure Bias for Diffusion Models}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  url = {https://openreview.net/forum?id=msUyy51yQe},
}
Figure