# CogView3: Finer and Faster Text-to-Image Generation via Relay Diffusion

**Authors:** Wendi Zheng, Jiayan Teng, Zhuoyi Yang, Weihan Wang, Jidong Chen, Xiaotao Gu, Yuxiao Dong, Ming Ding, Jie Tang  
**arXiv:** [2403.05121](https://arxiv.org/abs/2403.05121)  
**Date:** 2024-03-08  
**Summary page:** https://www.summarizepaper.com/en/paper/2403.05121v1/

## Key points

- Recent advancements in text-to-image generative systems are driven by diffusion models.
- Single-stage text-to-image diffusion models face challenges with computational efficiency and image detail refinement.
- CogView3 is a novel cascaded framework that implements relay diffusion in text-to-image generation.
- CogView3 creates low-resolution images initially and then applies relay-based super-resolution for enhanced output quality.
- CogView3 reduces both training and inference costs significantly compared to SDXL, the current state-of-the-art open-source text-to-image diffusion model.
- In human evaluations, CogView3 outperforms SDXL by 77.0% while requiring only half of the inference time.
- A distilled variant of CogView3 achieves comparable performance with just one-tenth of the inference time needed by SDXL.
- The study titled "CogView3: Finer and Faster Text-to-Image Generation via Relay Diffusion" presents a groundbreaking advancement in text-to-image generation promising enhanced efficiency and output quality.

## Summary

In the realm of text-to-image generative systems, recent advancements have been predominantly driven by diffusion models. However, single-stage text-to-image diffusion models encounter challenges related to computational efficiency and the refinement of image details. To address these issues, a novel cascaded framework called CogView3 has been introduced. stands out as the first model to implement relay diffusion in text-to-image generation, employing a unique approach of creating low-resolution images initially and then applying relay-based super-resolution for enhanced output quality. This innovative methodology not only yields competitive text-to-image results but also significantly reduces both training and inference costs. The experimental findings showcase the superior performance of  compared to SDXL, the current state-of-the-art open-source text-to-image diffusion model. In human evaluations,  surpasses SDXL by an impressive 77.0%, while requiring only half of the inference time. Furthermore, a distilled variant of  achieves comparable performance with just one-tenth of the inference time needed by SDXL. Authored by Wendi Zheng, Jiayan Teng, Zhuoyi Yang, Weihan Wang, Jidong Chen, Xiaotao Gu,Yuxiao Dong,Ming Ding,and Jie Tang,the study titled "CogView3: Finer and Faster Text-to-Image Generation via Relay Diffusion" presents a groundbreaking advancement in text-to-image generation that promises enhanced efficiency and output quality in this rapidly evolving field.