High-Resolution Image Synthesis with Latent Diffusion Models¶
Why this mattered¶
Rombach et al. made diffusion-based image generation substantially more practical by moving the denoising process from pixel space into the compressed latent space of a pretrained autoencoder. This was not merely an efficiency trick: the paper showed that a carefully chosen latent representation could preserve enough perceptual detail for high-quality synthesis while reducing the computational burden that had made pixel-space diffusion models expensive to train and slow to sample. In doing so, it helped shift diffusion models from impressive but resource-intensive research systems toward a general-purpose architecture for high-resolution image generation.
The second major contribution was architectural: cross-attention made the diffusion model flexibly conditionable on inputs such as text, segmentation maps, layouts, or bounding boxes. That design helped turn diffusion models into controllable image generators rather than only unconditional or narrowly class-conditional samplers. The paper therefore connected two trends that would define the next phase of generative modeling: scalable latent-space generation and natural-language or structured control over visual synthesis.
Its importance is especially clear in hindsight because latent diffusion became the technical basis for Stable Diffusion and a broader ecosystem of open text-to-image, inpainting, editing, and personalization systems. By reducing the cost of training and inference while preserving visual quality, the work changed who could build and modify high-resolution generative models: not only large industrial labs, but also academic groups, startups, and open-source communities. Subsequent breakthroughs in image editing, ControlNet-style conditioning, LoRA fine-tuning, and multimodal creative tools all built on the possibility that this paper made credible: powerful diffusion generation could be both high-resolution and widely deployable.
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 new state of the art scores for image inpainting and class-conditional image synthesis and highly competitive performance on various tasks, including unconditional image generation, text-to-image synthesis, and super-resolution, while significantly reducing computational requirements compared to pixel-based DMs.
Related¶
- cite → ImageNet: A large-scale hierarchical image database — Latent diffusion models use ImageNet as a large-scale benchmark dataset for evaluating class-conditional image synthesis.
- enables ← ImageNet: A large-scale hierarchical image database — ImageNet provided the large-scale natural-image benchmark and pretrained recognition backbone ecosystem that latent diffusion models used for high-resolution image synthesis evaluation and conditioning.