Rectified Diffusion: Straightness Is Not Your Need in Rectified Flow

Rectified Diffusion enhances generation speed and quality across various diffusion models.

First-order ODE focus unlocks consistent high-quality generation with fewer steps.

https://arxiv.org/abs/2410.07303

Original Problem 🔍:

Diffusion models excel in visual generation but suffer from slow generation speed due to computationally intensive ODE solving.

-----

Solution in this Paper 🛠️:

• Proposes Rectified Diffusion, generalizing rectified flow to broader diffusion models

• Keeps pretrained diffusion models unchanged, only using pre-collected noise-sample pairs for training

• Introduces Rectified Diffusion (Phased) for improved efficiency

• Incorporates consistency distillation for enhanced low-step performance

-----

Key Insights from this Paper 💡:

• Rectification's success lies in using pretrained models for matched noise-sample pairs

• First-order approximate ODE path is the essential training target, not straightness

• Rectified Diffusion simplifies training and extends to various diffusion forms

• First-order ODE supports consistent generation with arbitrary inference steps

-----

Results 📊:

• Outperforms InstaFlow with 8% of trained images for one-step generation

• Achieves better performance at 3% GPU days than InstaFlow's distilled model

• Consistently surpasses rectified flow-based methods across FID and CLIP scores

• Demonstrates superior human preference metrics compared to Rectified Flow