# InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation

**Authors:** Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, Qiang Liu  
**arXiv:** [2309.06380](https://arxiv.org/abs/2309.06380)  
**Date:** 2023-09-12  
**Summary page:** https://www.summarizepaper.com/en/paper/2309.06380v2/

## Key points

- Diffusion models have revolutionized text-to-image generation with their exceptional quality and creativity.
- Previous attempts to improve sampling speed and reduce computational costs through distillation have not resulted in a functional one-step model.
- The paper "InstaFlow" introduces Rectified Flow, including the innovative reflow procedure, to transform Stable Diffusion into an ultra-fast one-step model for text-to-image generation.
- InstaFlow achieves remarkable image quality with an FID of $23.3$ on MS COCO 2017-5k dataset, surpassing the previous state-of-the-art technique by a significant margin.
- Leveraging an expanded network with 1.7B parameters further improves the FID score to $22.4$, showcasing both efficiency and effectiveness in high-quality image synthesis tasks.
- InstaFlow sets a new benchmark for speed in image generation tasks, achieving an outstanding FID of $13.1$ on MS COCO 2014-30k dataset in just $0.09$ seconds.
- Training InstaFlow only requires 199 A100 GPU days, making it powerful and cost-effective for practical implementation.
- Codes and pre-trained models for InstaFlow are available at \url{github.com/gnobitab/InstaFlow}, enabling further exploration and replication of results.

## Summary

, , , , 

In the realm of text-to-image generation, diffusion models have emerged as a game-changer due to their remarkable quality and creativity. However, the multi-step sampling process inherent in these models has been a bottleneck, often necessitating numerous inference steps to achieve satisfactory results. Previous efforts to enhance sampling speed and reduce computational costs through distillation have fallen short in producing a functional one-step model. In this groundbreaking paper titled "InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation," authors Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, and Qiang Liu delve into the exploration of Rectified Flow, a recent method that has primarily been applied to small datasets. At the core of Rectified Flow lies the innovative \emph{reflow} procedure, which serves to straighten probability flow trajectories, refine noise-image coupling, and facilitate distillation with student models. The researchers propose a novel text-conditioned pipeline that transforms Stable Diffusion (SD) into an ultra-fast one-step model. Through their experimentation with reflow, they discover its pivotal role in enhancing the alignment between noise and images. Leveraging this new pipeline leads to the development of InstaFlow - the first one-step diffusion-based text-to-image generator capable of achieving SD-level image quality. Impressively, InstaFlow achieves an FID (Frechet Inception Distance) of $23.3$ on MS COCO 2017-5k dataset, surpassing the previous state-of-the-art technique known as progressive distillation by a significant margin ($37.2$ $\rightarrow$ $23.3$ in FID). By utilizing an expanded network with 1.7B parameters, the researchers further improve the FID score to $22.4$. Notably, on MS COCO 2014-30k dataset, InstaFlow achieves an outstanding FID of $13.1$ in just $0.09$ seconds - setting a new benchmark for speed in image generation tasks within $\leq 0.1$ second regime. This remarkable performance outshines recent approaches like StyleGAN-T ($13.9$ in $0.1$ second), showcasing the efficiency and effectiveness of InstaFlow in high-quality image synthesis tasks while maintaining rapid processing speeds. It is worth mentioning that training InstaFlow only requires 199 A100 GPU days - making it not only powerful but also cost-effective for practical implementation. For those interested in exploring further or replicating these results, codes and pre-trained models are readily available at \url{github.com/gnobitab/InstaFlow}. This research presented at ICLR 2024 marks a significant advancement in text-to-image generation technology and sets a new standard for high-quality image synthesis with unparalleled speed and efficiency.