Diffusion Models in Vision: A Survey

Analysis

Why This Paper Matters

This survey arrives at a pivotal moment in generative AI, where diffusion models have rapidly overtaken GANs as the state-of-the-art in image synthesis. With over 2,200 citations, it has become a canonical reference for anyone entering the field. The paper's value lies not just in cataloging methods but in providing a unified framework that connects three seemingly distinct approaches: DDPMs, score-based models, and SDEs. For practitioners, this clarity is essential for choosing the right formulation for their application.

The survey also bridges the gap between diffusion models and classical generative architectures, helping researchers understand where diffusion fits in the broader landscape. By explicitly comparing diffusion to VAEs, GANs, energy-based models, autoregressive models, and normalizing flows, the authors give readers a roadmap for hybrid approaches that could combine the strengths of multiple paradigms.

Technical Contributions

The paper's primary technical contribution is its systematic categorization of diffusion models into three core frameworks:

• Denoising Diffusion Probabilistic Models (DDPMs): The discrete-time formulation where noise is added and removed over a fixed number of steps.
• Noise Conditioned Score Networks (NCSNs): A continuous approach that learns the gradient of the data distribution's log-density (score function) at multiple noise levels.
• Stochastic Differential Equations (SDEs): A unifying framework that generalizes both DDPMs and NCSNs by modeling the diffusion process as a continuous-time SDE.

The survey further provides a multi-perspective categorization of vision applications, organizing papers by task (e.g., image generation, super-resolution, inpainting, segmentation) and by methodological innovation (e.g., sampling acceleration, conditional generation, latent space modeling). This dual taxonomy is invaluable for quickly locating relevant work.

Results

As a survey paper, the primary results are taxonomic and comparative rather than experimental. The authors synthesize findings from hundreds of papers to conclude that diffusion models consistently produce higher quality and more diverse samples than GANs on standard benchmarks like ImageNet and LSUN, though at the cost of slower sampling. Key comparisons include:

• Diffusion models achieve FID scores below 10 on ImageNet 256x256, outperforming GANs.
• Sampling typically requires 100-1000 steps, compared to GANs' single forward pass.
• Recent work on accelerated sampling (e.g., DDIM, DPM-solver) reduces steps to 10-50 with minimal quality loss.

The survey also notes that diffusion models excel in conditional generation tasks, particularly text-to-image synthesis (e.g., DALL·E 2, Imagen, Stable Diffusion), where they have become the dominant approach.

Significance

This survey has had a profound impact on the AI field by democratizing understanding of diffusion models. It has accelerated research by providing a clear taxonomy, enabling newcomers to quickly grasp the landscape. The paper's influence extends beyond vision into audio, video, and 3D generation, as the frameworks it describes are domain-agnostic. For Neura Market's audience of AI practitioners, this survey is essential reading for understanding the theoretical underpinnings of the models powering today's most impressive generative applications, from Stable Diffusion to Sora. Its identification of limitations—particularly sampling speed and computational cost—has directly motivated a wave of research on efficient sampling and model compression.