High-Resolution Image Synthesis with Latent Diffusion Models (2112.10752v2)

Abstract: By decomposing the image formation process into a sequential application of denoising autoencoders, diffusion models (DMs) achieve state-of-the-art synthesis results on image data and beyond. Additionally, their formulation allows for a guiding mechanism to control the image generation process without retraining. However, since these models typically operate directly in pixel space, optimization of powerful DMs often consumes hundreds of GPU days and inference is expensive due to sequential evaluations. To enable DM training on limited computational resources while retaining their quality and flexibility, we apply them in the latent space of powerful pretrained autoencoders. In contrast to previous work, training diffusion models on such a representation allows for the first time to reach a near-optimal point between complexity reduction and detail preservation, greatly boosting visual fidelity. By introducing cross-attention layers into the model architecture, we turn diffusion models into powerful and flexible generators for general conditioning inputs such as text or bounding boxes and high-resolution synthesis becomes possible in a convolutional manner. Our latent diffusion models (LDMs) achieve a new state of the art for image inpainting and highly competitive performance on various tasks, including unconditional image generation, semantic scene synthesis, and super-resolution, while significantly reducing computational requirements compared to pixel-based DMs. Code is available at https://github.com/CompVis/latent-diffusion .

Advancements in High-Resolution Image Synthesis

Introduction to Latent Diffusion Models

The computer vision community is constantly pushing the boundaries of what's possible with image synthesis. Recent approaches in the form of diffusion models have achieved impressive results in generating high-fidelity images. These models iteratively refine noise into detailed images through a reverse Markov process. While promising, such models often operate directly on pixels, making the optimization computationally intensive and the inference process time-consuming.

Optimization and Inference Efficiency

To address the computational challenges of traditional diffusion models, a novel approach applies these models in the latent space of autoencoders. Contrary to pixel space operation, latent diffusion models (LDMs) harness the efficiency of working with lower-dimensional representations. The use of autoencoders allows these models to reach a sweet spot between complexity reduction and detail preservation, thereby significantly reducing computation without sacrificing image quality.

Furthermore, the integration of cross-attention layers transforms diffusion models into potent generators. They can handle diverse conditioning inputs such as textual descriptions or bounding boxes, enabling high-resolution image synthesis through a more straightforward convolutional process.

Improvements in Image Synthesis Tasks

LDMs have shown remarkable versatility and performance across an array of image synthesis tasks. They have established new state-of-the-art benchmarks for image inpainting, class-conditional image synthesis, and demonstrated strong capabilities in tasks such as text-to-image synthesis and super-resolution. All the while, they've managed to substantially lower computational demands compared to pixel-based diffusion models.

Realizing High-Quality Conditional Generation

The LDMs stand out with their proficiency for conditional generation. With cross-attention mechanisms, diffusion models can seamlessly assimilate guidance from multimodal inputs. From incorporating class labels, dealing with masked image regions, or interpreting text descriptions, LDMs prove adept at a variety of conditional synthesis applications.

In the landscape of class-conditional image generation, these models provide high-quality outputs while utilizing fewer parameters and less computational power than leading alternatives. The method shows a practical pathway of achieving high-quality image synthesis with less environmental impact due to reduced energy consumption.

Conclusion

The research on LDMs breaks new ground by enhancing both the efficiency and feasibility of training and using diffusion models for image synthesis. From cutting down the need for extensive resources to democratizing the exploration of such models, this development is set to steer future research while tending to the pressing concern of increasing computational demands in AI. As these models become more accessible, they're poised to revolutionize image generation and editing, carving out possibilities across creative and commercial domains.