# Hyper-SD: Trajectory Segmented Consistency Model for Efficient Image Synthesis

**Authors:** Yuxi Ren, Xin Xia, Yanzuo Lu, Jiacheng Zhang, Jie Wu, Pan Xie, Xing Wang, Xuefeng Xiao  
**arXiv:** [2404.13686](https://arxiv.org/abs/2404.13686)  
**Date:** 2024-04-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2404.13686v1/

## Key points

- Growing interest in diffusion-aware distillation algorithms for reducing computational burden in Diffusion Models (DMs)
- Focus on ODE Trajectory Preservation and ODE Trajectory Reformulation
- Challenges of performance degradation and domain shifts in existing approaches
- Introduction of Hyper-SD framework combining strengths of both preservation and reformulation with near-lossless performance
- Trajectory Segmented Consistency Distillation for consistent distillation within predefined time-step segments
- Incorporation of human feedback learning to enhance model performance in low-step scenarios
- Integration of score distillation to improve low-step generation capabilities and unified LoRA support for inference at all steps
- State-of-the-art performance demonstrated across 1 to 8 inference steps for SDXL and SD1.5 models, with Hyper-SDXL outperforming SDXL-Lightning by +0.68 in CLIP Score and +0.51 in Aes Score during 1-step inference
- Authors: Yuxi Ren, Xin Xia, Yanzuo Lu, Jiacheng Zhang, Jie Wu, Pan Xie, Xing Wang, Xuefeng Xiao

## Summary

In recent years, there has been a growing interest in diffusion-aware distillation algorithms aimed at reducing the computational burden of multi-step inference processes in Diffusion Models (DMs). These algorithms typically focus on two key aspects: ODE Trajectory Preservation and ODE Trajectory Reformulation. However, existing approaches often suffer from performance degradation or domain shifts. To address these challenges, a novel framework called Hyper-SD has been proposed. Hyper-SD combines the strengths of both ODE Trajectory Preservation and Reformulation while maintaining near-lossless performance during step compression. The framework introduces Trajectory Segmented Consistency Distillation, which allows for consistent distillation within predefined time-step segments to preserve the original ODE trajectory from a higher-order perspective. Additionally, human feedback learning is incorporated to enhance model performance in low-step scenarios and mitigate any performance loss caused by the distillation process. Furthermore, Hyper-SD integrates score distillation to improve the model's low-step generation capabilities and introduces a unified LoRA to support the inference process at all steps. Extensive experiments and user studies have demonstrated that Hyper-SD achieves state-of-the-art performance across 1 to 8 inference steps for both SDXL and SD1.5 models. For instance, Hyper-SDXL outperforms SDXL-Lightning by +0.68 in CLIP Score and +0.51 in Aes Score during 1-step inference. The authors of this innovative framework include Yuxi Ren, Xin Xia, Yanzuo Lu, Jiacheng Zhang, Jie Wu, Pan Xie, Xing Wang, and Xuefeng Xiao. Their work titled "Hyper-SD: Trajectory Segmented Consistency Model for Efficient Image Synthesis" presents a significant advancement in addressing the limitations of current distillation techniques in diffusion models.