Visual Chain-of-Thought Diffusion Models

TL;DR

This paper introduces a two-stage sampling method for diffusion models that significantly improves unconditional image generation by leveraging conditional diffusion techniques, reducing FID scores substantially.

Contribution

It proposes a novel two-stage sampling approach that bridges the performance gap between conditional and unconditional diffusion models.

Findings

Improves FID by 25-50% over standard unconditional models

Leverages conditional diffusion models for unconditional generation

Demonstrates effectiveness across various image generation tasks

Abstract

Recent progress with conditional image diffusion models has been stunning, and this holds true whether we are speaking about models conditioned on a text description, a scene layout, or a sketch. Unconditional image diffusion models are also improving but lag behind, as do diffusion models which are conditioned on lower-dimensional features like class labels. We propose to close the gap between conditional and unconditional models using a two-stage sampling procedure. In the first stage we sample an embedding describing the semantic content of the image. In the second stage we sample the image conditioned on this embedding and then discard the embedding. Doing so lets us leverage the power of conditional diffusion models on the unconditional generation task, which we show improves FID by 25-50% compared to standard unconditional generation.