Rectified Diffusion: Straightness Is Not Your Need in Rectified Flow
Fu-Yun Wang, Ling Yang, Zhaoyang Huang, Mengdi Wang, Hongsheng Li
TL;DR
This paper introduces Rectified Diffusion, a simplified and more effective approach to rectified flow in diffusion models, emphasizing the importance of matched noise-sample pairs over straightness of the ODE path, and broadening applicability beyond flow-matching models.
Contribution
It proposes a new rectified diffusion method that simplifies training, removes unnecessary components, and extends rectification to all diffusion models, not just flow-matching ones.
Findings
Achieves superior performance on Stable Diffusion models.
Reduces training complexity and cost.
Validates effectiveness across multiple diffusion models.
Abstract
Diffusion models have greatly improved visual generation but are hindered by slow generation speed due to the computationally intensive nature of solving generative ODEs. Rectified flow, a widely recognized solution, improves generation speed by straightening the ODE path. Its key components include: 1) using the diffusion form of flow-matching, 2) employing -prediction, and 3) performing rectification (a.k.a. reflow). In this paper, we argue that the success of rectification primarily lies in using a pretrained diffusion model to obtain matched pairs of noise and samples, followed by retraining with these matched noise-sample pairs. Based on this, components 1) and 2) are unnecessary. Furthermore, we highlight that straightness is not an essential training target for rectification; rather, it is a specific case of flow-matching models. The more critical training target…
Peer Reviews
Decision·ICLR 2025 Poster
1. The authors generalized rectified flow to a wider array of diffusion models by focusing on paired noise-sample training, making the approach compatible with different prediction types and diffusion models. 2. The paper demonstrates that Rectified Diffusion performs better than prior methods like InstaFlow and Rectified Flow in terms of FID and CLIP scores, particularly when generating images in low-step regimes. 3. By dividing the ODE path into phases with enforced linearity in each phase, t
1. While the approach seems efficient, the paper builds on existing methods, particularly using paired noise-sample training, phased ODE segmentation and consistency distillation. The lack of entirely new techniques could limit its novelty in diffusion model research. 2. While this method employs distillation techniques that improve inference speed, achieving single-step image generation without pre-trained models could introduce a novel advancement. 3. The performance metrics on FID shows only
First strength is that the paper shows that straightness is necessary as required in rectified flow. It shows that rectified flow is a special case of rectified diffusion, which is what the paper proposed as a better generative model. Second strength is the better quantitative results and favorable comparison with other methods.
Maybe I am missing something, it seems the paper argues that straightness of the ODE trajectory is not necessary, but its results appear to show a straight path, such as Figure 4 and Figure 5. Other weaknesses include that from visual evaluation, it seems that though the rectified diffusion gave better images than rectified flow in most cases, the improvement is not always clear. For example, in Figure 10, the result of rectified diffusion on the sneakers seem that the string missed a buttonhol
They tried to highlight that straightness is not an essential training target for rectification.
It was not stated clearly why it is necessary to study whether straightness is not an essential training target for rectification. To me, it was the straightness used in the ode of the rectified flow algorithm that causes the usage of rectified flow, thus it seems trivial that there is no need to use the rectified flow in a ode that does not follow a straight line. The diffusion that follows the curve form that the authors proposed is mentioned in some papers including the original rectified f
Code & Models
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsFluid Dynamics and Turbulent Flows · Rheology and Fluid Dynamics Studies
MethodsSPEED: Separable Pyramidal Pooling EncodEr-Decoder for Real-Time Monocular Depth Estimation on Low-Resource Settings · Diffusion