Unifying Diffusion Models' Latent Space, with Applications to CycleDiffusion and Guidance (2210.05559v2)

Abstract: Diffusion models have achieved unprecedented performance in generative modeling. The commonly-adopted formulation of the latent code of diffusion models is a sequence of gradually denoised samples, as opposed to the simpler (e.g., Gaussian) latent space of GANs, VAEs, and normalizing flows. This paper provides an alternative, Gaussian formulation of the latent space of various diffusion models, as well as an invertible DPM-Encoder that maps images into the latent space. While our formulation is purely based on the definition of diffusion models, we demonstrate several intriguing consequences. (1) Empirically, we observe that a common latent space emerges from two diffusion models trained independently on related domains. In light of this finding, we propose CycleDiffusion, which uses DPM-Encoder for unpaired image-to-image translation. Furthermore, applying CycleDiffusion to text-to-image diffusion models, we show that large-scale text-to-image diffusion models can be used as zero-shot image-to-image editors. (2) One can guide pre-trained diffusion models and GANs by controlling the latent codes in a unified, plug-and-play formulation based on energy-based models. Using the CLIP model and a face recognition model as guidance, we demonstrate that diffusion models have better coverage of low-density sub-populations and individuals than GANs. The code is publicly available at https://github.com/ChenWu98/cycle-diffusion.

Overview of "Unifying Diffusion Models' Latent Space"

The paper "Unifying Diffusion Models' Latent Space, with Applications to CycleDiffusion and Guidance" by Chen Henry Wu and Fernando De la Torre addresses the ongoing evolution in generative modeling through diffusion models. Traditionally, diffusion models employ a sequence of denoised samples as their latent representations, unlike the straightforward Gaussian latent spaces used in GANs, VAEs, and normalizing flows. This work proposes an innovative Gaussian-based latent space formulation for diffusion models, offering a novel perspective on their latent structure.

Key Contributions

1. Gaussian Latent Space for Diffusion Models: The authors reframe the latent structure of diffusion models into a Gaussian latent space. This approach parallels traditional generative models, enabling deterministic mappings from Gaussian noise to images akin to GANs, VAEs, and normalizing flows.
2. DPM-Encoder: Introducing the DPM-Encoder, a reconstructable encoder for stochastic diffusion probabilistic models (DPMs), the paper resolves the challenge of encoding real images into latent spaces for these models. This ensures deterministic reconstruction, a notable advantage which aids in image to latent code mappings.
3. CycleDiffusion: Utilizing the common latent space framework, the CycleDiffusion method allows unpaired image-to-image translation. This is based on the observation that independent diffusion models trained on related domains can produce similar images from fixed latent codes. CycleDiffusion is also extended to perform zero-shot image-to-image editing using text-to-image diffusion models.
4. Plug-and-Play Guidance: The research further extends the capability of diffusion models to allow plug-and-play guidance using energy-based models. This unifies the approach to guiding pre-trained GANs and diffusion models without fine-tuning on noisy images. It demonstrates that diffusion models can effectively cover low-density sub-populations better than GANs.

Results and Implications

The empirical results highlight significant performance improvements with CycleDiffusion, surpassing prior GAN and diffusion-based methods in unpaired image-to-image translation tasks. It is also shown that diffusion models can manipulate images in a zero-shot setting, emphasizing their adaptability and utility in real-world tasks without additional training.

The improvements in guidance methods presented allow for more granular control over generated content, suggesting broader applications for image editing and synthesis that are robust against the diversity of real-world data distributions.

Future Directions

The paper opens avenues for extending theoretical understanding of the latent spaces in stochastic and deterministic diffusion models. Further exploration could integrate insights from optimal transport theory, as mentioned in related works. Also, improving the efficiency of high-resolution image guidance presents a promising research direction.

In conclusion, the unification of diffusion models' latent spaces proposed in this paper represents a crucial step in bridging the gap between different generative modeling paradigms. This work not only offers technical advancements but also lays the groundwork for more integrated and flexible generative models in artificial intelligence research.