Latent Diffusion Models Achieve High-Resolution Image Synthesis with Less Computing Power
Researchers introduce Latent Diffusion Models that operate in compressed latent space rather than pixel space, drastically cutting computational costs while maintaining top-tier image quality. This breakthrough enables flexible, high-resolution image generation guided by text or bounding boxes.
Diffusion models achieve remarkable image synthesis results by applying denoising autoencoders sequentially, but they traditionally operate directly in pixel space. This standard approach consumes hundreds of GPU days for training and makes inference expensive due to sequential evaluations. To solve this inefficiency, researchers now apply diffusion models within the latent space of powerful pretrained autoencoders rather than raw pixels.
This new method, known as Latent Diffusion Models (LDMs), strikes a near-optimal balance between complexity reduction and detail preservation. By working in a compressed representation, the system greatly boosts visual fidelity while significantly reducing computational requirements. The architecture also incorporates cross-attention layers to handle general conditioning inputs like text prompts or bounding boxes.
LDMs establish a new state of the art for image inpainting and deliver highly competitive performance on unconditional image generation, semantic scene synthesis, and super-resolution tasks. This approach allows high-resolution synthesis to happen in a convolutional manner, making advanced image generation accessible on limited computational resources without sacrificing quality or flexibility.