From Diffusion to One-Step Generation: A Comparative Study of Flow-Based Models with Application to Image Inpainting

TL;DR

This paper compares diffusion, flow-based, and one-step generative models for images, showing that flow models like MeanFlow can generate high-quality images faster and effectively perform image inpainting.

Contribution

It introduces a unified implementation of three generative models, demonstrating the efficiency of MeanFlow for single-step generation and extending CFM to image inpainting with significant quality improvements.

Findings

CFM achieves FID 24.15 with 50 steps, outperforming DDPM.

MeanFlow attains FID 29.15 with one-step sampling, 50X faster.

Inpainting with CFM improves PSNR and SSIM significantly.

Abstract

We present a comprehensive comparative study of three generative modeling paradigms: Denoising Diffusion Probabilistic Models (DDPM), Conditional Flow Matching (CFM), and MeanFlow. While DDPM and CFM require iterative sampling, MeanFlow enables direct one-step generation by modeling the average velocity over time intervals. We implement all three methods using a unified TinyUNet architecture (<1.5M parameters) on CIFAR-10, demonstrating that CFM achieves an FID of 24.15 with 50 steps, significantly outperforming DDPM (FID 402.98). MeanFlow achieves FID 29.15 with single-step sampling -- a 50X reduction in inference time. We further extend CFM to image inpainting, implementing mask-guided sampling with four mask types (center, random bbox, irregular, half). Our fine-tuned inpainting model achieves substantial improvements: PSNR increases from 4.95 to 8.57 dB on center masks (+73%), and SSIM improves from 0.289 to 0.418 (+45%), demonstrating the effectiveness of inpainting-aware training.