Abstract:Diffusion Models have demonstrated remarkable performance in image generation. However, their demanding computational requirements for training have prompted ongoing efforts to enhance the quality of generated images through modifications in the sampling process. A recent approach, known as Discriminator Guidance, seeks to bridge the gap between the model score and the data score by incorporating an auxiliary term, derived from a discriminator network. We show that despite significantly improving sample quality, this technique has not resolved the persistent issue of Exposure Bias and we propose SEDM-G++, which incorporates a modified sampling approach, combining Discriminator Guidance and Epsilon Scaling. Our proposed approach outperforms the current state-of-the-art, by achieving an FID score of 1.73 on the unconditional CIFAR-10 dataset.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the Exposure Bias problem in Diffusion Models when generating images. Although the introduction of Discriminator Guidance technology significantly improves the sample quality, it fails to effectively alleviate the Exposure Bias problem. Specifically: 1. **Definition of Exposure Bias**: Exposure Bias refers to the difference between the input data in the training stage and the inference stage. During the training process, the model can access the real training samples \(x_t\), while during the inference process, the model can only rely on the previously generated samples \(\hat{x}_t\). This difference leads to the accumulation of noise prediction errors, thus affecting the quality of the generated samples. 2. **Limitations of Existing Methods**: - **Discriminator Guidance (DG)**: By introducing a discriminative network to adjust the gap between the model score and the real data score, although it improves the sample quality, it fails to solve the Exposure Bias problem. - **Other Methods**: Methods such as Epsilon Scaling have been proposed to reduce Exposure Bias, but these methods have their own advantages and disadvantages and are more difficult to implement. 3. **Solution Proposed in the Paper**: The authors propose SEDM - G++, an improved sampling method that combines Discriminator Guidance and Epsilon Scaling. In this way, SEDM - G++ not only improves the sample quality but also effectively reduces the Exposure Bias. ### Specific Problem Description - **Influence of Exposure Bias**: In Diffusion Models, because the samples used in the inference process are different from those in the training process, it leads to the accumulation of noise prediction errors, which in turn affects the quality of the generated samples. - **Limitations of Discriminator Guidance**: Although Discriminator Guidance significantly improves the sample quality, it fails to effectively solve the Exposure Bias problem. Especially when using the Euler solver, the sample quality actually decreases. - **Role of Epsilon Scaling**: Epsilon Scaling reduces Exposure Bias by scaling the predicted noise factor and is a plug - in method without the need to retrain the model. ### Contributions of the Paper - **Research on Exposure Bias**: Analyzed the influence of Discriminator Guidance on Exposure Bias in Diffusion Models. - **Proposing SEDM - G++**: Combined Discriminator Guidance and Epsilon Scaling and proposed a new sampling method. - **Experimental Verification**: Conducted extensive experiments on the CIFAR - 10 dataset and proved the effectiveness of SEDM - G++, achieving the best result with an FID score of 1.73. In conclusion, this paper aims to solve the Exposure Bias problem in Diffusion Models when generating images, and by proposing the SEDM - G++ method, it achieves better sample quality and less Exposure Bias.

Mitigating Exposure Bias in Discriminator Guided Diffusion Models

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

Elucidating the Exposure Bias in Diffusion Models

InvDiff: Invariant Guidance for Bias Mitigation in Diffusion Models

Refining Generative Process with Discriminator Guidance in Score-based Diffusion Models

Measurement Guidance in Diffusion Models: Insight from Medical Image Synthesis

Guiding a Diffusion Model with a Bad Version of Itself

Exploring Guided Sampling of Conditional GANs

Enhancing Diffusion-Based Image Synthesis with Robust Classifier Guidance

Gradient-Free Classifier Guidance for Diffusion Model Sampling

AdaDiff: Accelerating Diffusion Models through Step-Wise Adaptive Computation

Elucidating the Design Space of Diffusion-Based Generative Models

Improving Sample Quality of Diffusion Models Using Self-Attention Guidance

Segmentation-Free Guidance for Text-to-Image Diffusion Models

E2EDiff: Direct Mapping from Noise to Data for Enhanced Diffusion Models

Unmasking Bias in Diffusion Model Training

Bridging the Gap: Addressing Discrepancies in Diffusion Model Training for Classifier-Free Guidance

An Efficient Framework for Enhancing Discriminative Models via Diffusion Techniques

Simpler Diffusion (SiD2): 1.5 FID on ImageNet512 with pixel-space diffusion

Reducing Spatial Fitting Error in Distillation of Denoising Diffusion Models

Elucidating The Design Space of Classifier-Guided Diffusion Generation