High-Resolution Image Synthesis with Latent Diffusion Models

Abstract

By decomposing the image formation process into a sequential application of denoising autoencoders, diffusion models (DMs) achieve state-of-the-art synthesis results. We apply diffusion models in the latent space of powerful pretrained autoencoders, achieving a near-optimal point between complexity reduction and detail preservation, greatly boosting visual fidelity.

Key Findings

• 1Introduced Latent Diffusion Models (LDMs) operating in compressed latent space
• 2Achieved high-quality image synthesis with dramatically reduced computational cost
• 3Enabled text-to-image generation accessible on consumer hardware
• 4Demonstrated cross-attention conditioning for flexible guided synthesis
• 5Achieved state-of-the-art results on image inpainting, super-resolution, and generation

Impact & Significance