landmark

High-Resolution Image Synthesis with Latent Diffusion Models

Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, Björn Ommer · LMU Munich, Runway, Heidelberg University · 2021-12 · arXiv:2112.10752 · code

Why it matters

Moves diffusion training from pixel space into the latent space of a pretrained autoencoder, cutting compute by orders of magnitude while keeping quality; introduces the VAE + UNet + cross-attention recipe every 'latent diffusion' system since has started from.

What this paper does

read: full text

Before this

Diffusion models produced state-of-the-art image samples but operated directly on pixels. Training the strongest pixel-space models took 150 to 1000 V100 GPU-days, and producing 50,000 samples took roughly 5 days on a single A100, putting them out of reach for most researchers.

The problem

Most bits in a digital image are imperceptible detail. A diffusion model trained directly on pixels still has to evaluate its network and compute gradients over all of those bits, so compute is spent on information a viewer never notices.

The idea

Separate the generative model from the pixel-level compression it does not need. First train an autoencoder that removes imperceptible high-frequency detail while preserving the true image manifold, then run the diffusion process entirely inside that compact latent space, adding cross-attention layers so the same architecture can be conditioned on text, layouts, or other signals.

How it works

An encoder E downsamples an image by a factor f=2^m and a decoder D reconstructs it, trained with a perceptual loss plus a patch-based adversarial objective and either a mild KL penalty toward a standard normal (KL-regularization, close to a VAE) or a vector-quantization layer in the decoder (VQ-regularization). A time-conditional UNet then denoises latents z_t with the standard noise-prediction objective and never touches pixels directly; E and D are frozen once trained, so decoding a sample needs one forward pass through D. Conditioning enters through added cross-attention layers, where a domain-specific encoder tau_theta maps input y, such as text or a class label, to an intermediate representation that supplies the keys and values, while the UNet's own features supply the queries, so tau_theta and the UNet are trained jointly under one loss. Downsampling factor is an explicit design choice. f=1, pixel space, leaves an FID gap of 38 versus LDM-8 after 2M training steps on the same architecture, f=4 and f=8 give the best balance of compression and fidelity, and f=32 loses too much information.

Evidence

LDM-4 reaches FID 5.11 on CelebA-HQ, 4.98 on FFHQ, 4.02 on LSUN-Churches (LDM-8) and 2.95 on LSUN-Bedrooms, approaching pixel-space ADM's 1.90 on LSUN-Bedrooms at a fraction of the compute. In text-to-image synthesis LDM-4 reaches FID 17.01 with 645M parameters, ahead of VQGAN+T's 28.86 and ImageBART's 22.61 despite roughly half ImageBART's parameter count. On Places inpainting at 512x512 LDM-4 (big, fine-tuned) reaches FID 9.39 against LaMa's 12.0. Across tasks the paper reports at least a 2.7x training speed-up and at least 1.6x better FID relative to pixel-space diffusion at matched settings.

Limitations

The authors concede that sequential sampling remains slower than GANs even in latent space. They also concede that using an autoencoder is questionable when high precision is required, since although reconstruction loss is small for their f=4 models, the autoencoder's reconstruction fidelity can become a bottleneck for tasks needing fine-grained pixel accuracy, such as super-resolution.

Why it matters

This is the VAE-plus-UNet-plus-cross-attention recipe that Stable Diffusion shipped, and that essentially every later latent diffusion system, including DiT and SD3 further down this lineage, has started from. Moving the expensive iterative process out of pixel space cut training cost from hundreds of GPU-days to a range within reach of ordinary university and startup compute budgets.

Abstract, in the authors' own words

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 .

Research line

VAE-latent diffusiondominant

Reported results

Benchmark	Value	Guidance	Budget	Source
ImageNet 256x256 rFID	0.62	—	—	MAETok Table 4

Design-axis choices

Representation

VAE latent

Architecture

UNet backbone

Conditioning & control

Cross-attention conditioning

Method note — the shared flow-matching interpolation

Every flow-matching / rectified-flow paper in this atlas trains toward a straight-line path between a noise sample x₀ and a data sample x₁:

x_t = (1-t)\,x_0 + t\,x_1, \qquad v_\theta(x_t, t) \approx x_1 - x_0

Builds on

Nothing recorded yet.

Built on by

Scalable Diffusion Models with Transformers builds_on this — DiT operates in LDM's VAE latent space, replacing only the denoising backbone (UNet -> ViT).
Scaling Rectified Flow Transformers for High-Resolution Image Synthesis improves this — SD3 is a direct production-scale successor in the 'latent diffusion' lineage LDM started, with a new architecture and objective.
Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models improves this — VA-VAE is a drop-in successor to SD-VAE-style tokenizers, aligned to a vision foundation model instead of trained for reconstruction alone.