Image Diffusion Preview with Consistency Solver
Fu-Yun Wang, Hao Zhou, Liangzhe Yuan, Sanghyun Woo, Boqing Gong, Bohyung Han, Ming-Hsuan Yang, Han Zhang, Yukun Zhu, Ting Liu, Long Zhao
2025-12-16
Summary
This paper introduces a new way to speed up image creation with diffusion models, focusing on making the process more interactive for users.
What's the problem?
Creating images with diffusion models can be slow, which makes it frustrating when you want to quickly experiment and refine an image idea. Existing methods to make these models faster either don't produce high-quality initial images or don't guarantee that the quick preview will look like the final, fully-rendered image.
What's the solution?
The researchers developed a technique called 'Diffusion Preview' that generates a rough draft of the image very quickly using a small number of processing steps. To make these quick drafts look good and consistent with the final image, they created a new tool called 'ConsistencySolver'. This tool is like a smart algorithm that learns to predict how the image should evolve, and it's optimized using a method called Reinforcement Learning. It's designed to be efficient and doesn't require a lot of extra training.
Why it matters?
This work is important because it significantly reduces the time it takes to create images with diffusion models, almost halving the amount of time a user spends interacting with the system. By providing a fast and reliable preview, users can quickly see if they're on the right track and avoid wasting time refining images that aren't what they want, ultimately making the image generation process much more user-friendly and efficient.
Abstract
The slow inference process of image diffusion models significantly degrades interactive user experiences. To address this, we introduce Diffusion Preview, a novel paradigm employing rapid, low-step sampling to generate preliminary outputs for user evaluation, deferring full-step refinement until the preview is deemed satisfactory. Existing acceleration methods, including training-free solvers and post-training distillation, struggle to deliver high-quality previews or ensure consistency between previews and final outputs. We propose ConsistencySolver derived from general linear multistep methods, a lightweight, trainable high-order solver optimized via Reinforcement Learning, that enhances preview quality and consistency. Experimental results demonstrate that ConsistencySolver significantly improves generation quality and consistency in low-step scenarios, making it ideal for efficient preview-and-refine workflows. Notably, it achieves FID scores on-par with Multistep DPM-Solver using 47% fewer steps, while outperforming distillation baselines. Furthermore, user studies indicate our approach reduces overall user interaction time by nearly 50% while maintaining generation quality. Code is available at https://github.com/G-U-N/consolver.