[ICLR2024] ⚡InstaFlow! One-Step Stable Diffusion with Rectified Flow

[Paper] [Demo in 🤗Hugging Face Space] [Code and Pre-trained Models][Colab Notebook]

by Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, Qiang Liu

News

• (New) 2024/06/07 Our large-scale Rectified Flow is extended to text-to-3D and image inversion/editing! Check out the amazing work from Xiaofeng Yang et al. (paper and code)!
• 2024/05/17 Try our new few-step model PeRFlow at here!
• 2023/12/04 We updated the demo in 🤗Hugging Face Space with InstaFlow+dreamshaper-7. Image quality significantly improves! We also provide the Gradio demo for you to run locally here.
• 2023/12/04 One-step InstaFlow is compatible with pre-trained LoRAs! See here. Code is available here. (We thank individual contributor Dr. Hanshu Yan)
• 2023/12/04 ONNX support is available now! [ONNX InstaFlow] [ONNX 2-Rectified Flow] [ONNXStack UI] (We thank saddam213)
• 2023/11/23 Colab notebook is online now. Try it here. (We thank individual contributor xaviviro)
• 2023/11/22 One-step InstaFlow is compatible with pre-trained ControlNets. See here. (We thank individual contributor Dr. Hanshu Yan)
• 2023/11/22 We release the pre-trained models and inference codes here.
• 2023/09/26 We provide a demo of InstaFlow-0.9B in 🤗Hugging Face Space. Try it here.

Introduction

Diffusion models have demonstrated remarkable promises in text-to-image generation. However, their efficacy is still largely hindered by computational constraints stemming from the need of iterative numerical solvers at the inference time for solving the diffusion/flow processes.

InstaFlow is an ultra-fast, one-step image generator that achieves image quality close to Stable Diffusion, significantly reducing the demand of computational resources. This efficiency is made possible through a recent Rectified Flow technique, which trains probability flows with straight trajectories, hence inherently requiring only a single step for fast inference.

InstaFlow has several advantages:

• Ultra-Fast Inference: InstaFlow models are one-step generators, which directly map noises to images and avoid multi-step sampling of diffusion models. On our machine with A100 GPU, the inference time is around 0.1 second, saving ~90% of the inference time compared to the original Stable Diffusion.
• High-Quality: InstaFlow generates images with intricate details like Stable Diffusion, and have similar FID on MS COCO 2014 as state-of-the-art text-to-image GANs, like StyleGAN-T.
• Simple and Efficient Training: The training process of InstaFlow merely involves supervised training. Leveraging pre-trained Stable Diffusion, it only takes 199 A100 GPU days to get InstaFlow-0.9B.

Gallery

One-step generation with InstaFlow-0.9B (0.09s per image, $512 \times 512$)

One-step generation with InstaFlow-1.7B (0.12s per image, $512 \times 512$)

One-step generation with InstaFlow-0.9B (0.09s) + SDXL-Refiner ($1024 \times 1024$)

Latent space interpolation of one-step InstaFlow-0.9B (0.09s per image, $512 \times 512$)

interpolation.mp4

LoRA

One-step InstaFlow is compatible with pre-trained LoRAs. We thank individual contributor Dr. Hanshu Yan for providing and testing the Rectified Flow+LoRA pipeline!

InstaFlow seems to have higher diversity than SDXL-Turbo.

lora.mp4

ControlNet

One-step InstaFlow is fully compatible with pre-trained ControlNets. We thank individual contributor Dr. Hanshu Yan for providing and testing the Rectified Flow+ControlNet pipeline!

Below are One-Step Generation with InstaFlow-0.9B + ControlNet:

Comparison with SD 1.5 on our A100 machine

For an intuitive understanding, we used the same A100 server and took screenshots from the Gridio interface of random generation with different models. InstaFlow-0.9B is one-step, while SD 1.5 adopts 25-step DPMSolver. It takes around 0.3 second to download the image from the server. The text prompt is "A photograph of a snowy mountain near a beautiful lake under sunshine."

InstaFlow-0.9B	Stable Diffusion 1.5

Method: Straightening Generative Probability Flows with Text-Conditioned Reflow

method_github.mov

Our pipeline consists of three steps:

1. Generate (text, noise, image) triplets from pre-trained Stable Diffusion
2. Apply text-conditioned reflow to yield 2-Rectified Flow, which is a straightened generative probability flow.
3. Distill from 2-Rectified Flow to get One-Step InstaFlow. Note that distillation and reflow are orthogonal techniques.

As captured in the video and the image, straight flows have the following advantages:

• Straight flows require fewer steps to simulate.
• Straight flows give better coupling between the noise distribution and the image distribution, thus allow successful distillation.

Related Materials

We provide several related links and readings here:

• The official Rectified Flow github repo (https://github.com/gnobitab/RectifiedFlow)

• An introduction of Rectified Flow (https://www.cs.utexas.edu/~lqiang/rectflow/html/intro.html)

• An introduction of Rectified Flow in Chinese--Zhihu (https://zhuanlan.zhihu.com/p/603740431)

• FlowGrad: Controlling the Output of Generative ODEs With Gradients (https://github.com/gnobitab/FlowGrad)

• Fast Point Cloud Generation with Straight Flows (https://github.com/klightz/PSF)

• Piecewise Rectified Flow (https://github.com/magic-research/piecewise-rectified-flow)

• Text-to-Image Rectified Flow as Plug-and-Play Priors (https://github.com/yangxiaofeng/rectified_flow_prior)

Citation

@inproceedings{liu2023instaflow,
  title={Instaflow: One step is enough for high-quality diffusion-based text-to-image generation},
  author={Liu, Xingchao and Zhang, Xiwen and Ma, Jianzhu and Peng, Jian and Liu, Qiang},
  booktitle={International Conference on Learning Representations},
  year={2024}
}

Thanks

Our training scripts are modified from one of the fine-tuning examples in Diffusers. Other parts of our work also heavily relies on the 🤗 Diffusers library.