Refining Generative Process with Discriminator Guidance in Score-based Diffusion Models (2211.17091v4)

Abstract: The proposed method, Discriminator Guidance, aims to improve sample generation of pre-trained diffusion models. The approach introduces a discriminator that gives explicit supervision to a denoising sample path whether it is realistic or not. Unlike GANs, our approach does not require joint training of score and discriminator networks. Instead, we train the discriminator after score training, making discriminator training stable and fast to converge. In sample generation, we add an auxiliary term to the pre-trained score to deceive the discriminator. This term corrects the model score to the data score at the optimal discriminator, which implies that the discriminator helps better score estimation in a complementary way. Using our algorithm, we achive state-of-the-art results on ImageNet 256x256 with FID 1.83 and recall 0.64, similar to the validation data's FID (1.68) and recall (0.66). We release the code at https://github.com/alsdudrla10/DG.

• The paper introduces Discriminator Guidance (DG) as a novel technique to refine the denoising process in score-based diffusion models.
• The method leverages a post-training discriminator to stabilize and accelerate convergence, yielding competitive FID and recall on ImageNet.
• DG offers practical benefits by reducing computational costs while effectively aligning generated samples with real-world data distributions.

An Overview of Refining Generative Processes with Discriminator Guidance in Score-based Diffusion Models

The paper "Refining Generative Process with Discriminator Guidance in Score-based Diffusion Models" proposes a novel approach to enhance the sample quality in pre-trained diffusion models. The authors introduce a technique called Discriminator Guidance (DG), which involves deploying a discriminator to offer explicit supervision during the denoising phase of a sample path. This method uses the pre-trained score model and introduces a new mechanism where the discriminator helps refine the generative process.

Methodology

Discriminator Guidance is distinct from Generative Adversarial Networks (GANs) as there is no requirement for joint training of score and discriminator networks. The discriminator is trained subsequently to the score model, which stabilizes and accelerates the convergence process. During sample generation, DG augments the pre-trained score with an auxiliary component that is designed to deceive the discriminator, correcting the model score to align with the optimal data score.

The authors achieve state-of-the-art (SOTA) results on the ImageNet 256x256 benchmark, recording a Fréchet Inception Distance (FID) of 1.83 and a recall of 0.64, performance metrics that are comparable to the validation data's FID of 1.68 and recall of 0.66. This significant achievement suggests that DG effectively guides the sample generation closer to real-world data distributions.

Theoretical Insights and Empirical Validation

The paper supports its methodology with theoretical foundations. It introduces a correction term designed to adjust the model score to match the data score at an optimal discriminator. This is backed up by rigorous experimentation, where the authors show that DG yields demonstrable improvements in image datasets such as CIFAR-10, CelebA, FFHQ 64x64, and ImageNet 256x256.

Implications and Future Directions

This work shines a light on the utility of using a discriminator as a complementary tool alongside pre-trained diffusion models. It opens up pathways for reducing computational costs by maintaining pre-trained scores fixed and using auxiliary components for further refinement. The paper suggests potential for DG to address large-time poor estimation issues identified in score matching frameworks, which could lead to improvements in other generative tasks.

The results indicate that adding discriminator guidance could become a pivotal strategy in enhancing the capabilities of existing pre-trained models, particularly in scenarios where fine-tuning or retraining entire models may not be feasible. Practically, this could mean more efficient computation without compromising on the diversity or quality of generated samples.

Looking forward, additional research could be directed towards combining DG with other guidance techniques to tackle mode collapse and other inherent limitations of pre-trained generative models. Moreover, understanding how DG interacts with different SDE frameworks could further improve model efficacy and broaden its applications.

Conclusion

In essence, the method proposed by the authors presents an innovative step in refining diffusion model outputs through discriminator guidance. By methodically aligning the generated samples closer to true data distributions, it extends the boundaries of what is achievable with score-based generative models. This research delivers practical insights and emphasizes the ongoing potential for refinement in generative models, paving the way for further exploration and development in the field of artificial intelligence.