Data

Tools

Indexes

Applications

Experiments

Awesome

An open API service indexing awesome lists of open source software.

https://github.com/lmxyy/sige

[NeurIPS 2022, T-PAMI 2023] Efficient Spatially Sparse Inference for Conditional GANs and Diffusion Models
https://github.com/lmxyy/sige

acceleration conditional-gan conditional-gans ddim ddpm diffusion-models gans gaugan image-editing progressive-distillation sparse sparse-convolution

Last synced: 14 days ago
JSON representation

[NeurIPS 2022, T-PAMI 2023] Efficient Spatially Sparse Inference for Conditional GANs and Diffusion Models

Awesome Lists containing this project

README 

        
# Spatially Incremental Generative Engine (SIGE)



### [Paper](https://arxiv.org/abs/2211.02048) | [Project](https://www.cs.cmu.edu/~sige/) | [Slides](https://www.cs.cmu.edu/~sige/resources/slides.key) | [YouTube](https://youtu.be/rDPotGoPPkQ)



**[NEW!]** SIGE is accepted by T-PAMI!



**[NEW!]** SIGE supports [Stable Diffusion](./stable_diffusion) and Mac MPS backend! We also release the codes of [an interactive demo for DDPM](diffusion_demo) on M1 Macbook Pro!



**[NEW!]** SIGE is accepted by NeurIPS 2022! Our code and benchmark datasets are publicly available!



![teaser](https://github.com/lmxyy/sige/raw/main/assets/teaser.jpg)

*We introduce Spatially Incremental Generative Engine (SIGE),an engine that selectively performs computations at the edited regions for image editing applications. The computation and latency are measured for a single forward. For the above examples, SIGE significantly reduces the computation of [SDEdit](https://github.com/ermongroup/SDEdit) with [DDPM](https://github.com/lucidrains/denoising-diffusion-pytorch) (4-6x) and [Stable Diffusion](https://github.com/CompVis/stable-diffusion) (8x), and [GauGAN](https://github.com/NVlabs/SPADE) (15x) while preserving the image quality. When combined with existing model compression methods such as [GAN Compression](https://github.com/mit-han-lab/gan-compression), it further reduces the computation of GauGAN by 47x. On NVIDIA RTX 3090, SIGE achieves up to 7.2x speedups.*



Efficient Spatially Sparse Inference for Conditional GANs and Diffusion Models

[Muyang Li](https://lmxyy.me/), [Ji Lin](http://linji.me/), [Chenlin Meng](https://cs.stanford.edu/~chenlin/), [Stefano Ermon](https://cs.stanford.edu/~ermon/), [Song Han](https://songhan.mit.edu/), and [Jun-Yan Zhu](https://www.cs.cmu.edu/~junyanz/)

CMU, MIT, and Stanford

In NeurIPS 2022.



## Demo





  




SIGE achieves 2x less conversion time compared to original DDPM on M1 MacBook Pro GPU as we selectively perform computation on the edited regions.



## Overview



![overview](https://github.com/lmxyy/sige/raw/main/assets/method.gif)*Tiling-based sparse convolution overview. For each convolution Fl in the network, we wrap it into SIGE Convl. The activations of the original image are already pre-computed. When getting the edited image, we first compute a difference mask between the original and edited image and reduce the mask to the active block indices to locate the edited regions. In each SIGE Convl, we directly gather the active blocks from the edited activation Aledited according to the reduced indices, stack the blocks along the batch dimension, and feed them into Fl. The gathered blocks have an overlap of width 2 if Fl is 3×3 convolution with stride 1. After getting the output blocks from Fl, we scatter them back into Fl(Aloriginal) to get the edited output, which approximates Fl(Aledited).*



## Performance



### Efficiency



![overview](https://github.com/lmxyy/sige/raw/main/assets/results.jpg)

*With 1.2% edits, SIGE could reduce the computation of DDPM, Progressive Distillation and GauGAN by 7-18x, achieve a 2-4x speedup on NVIDIA RTX 3090, 3-5x speedup on Apple M1 Pro GPU and 4-14x on M1 Pro CPU. When combined with GAN Compression, it further reduces 50x computation on GauGAN, achieving 38x speedup on M1 Pro CPU. Please check our paper for more details and results.*



### Quality



![overview](https://github.com/lmxyy/sige/raw/main/assets/quality.jpg)*Qualitative results under different edit sizes. PD is Progressive Distillation. Our method well preserves the visual fidelity of the original model without losing global context.*



![quality-stable-diffusion](https://github.com/lmxyy/sige/raw/main/assets/quality-stable-diffusion.jpg)



*More qualitative results of Stable Diffusion on both image inpainting and editing, measured on NVIDIA RTX 3090.*



References:



* Denoising Diffusion Probabilistic Models (DDPM), Ho et al., ICLR 2020

* Denoising Diffusion Implicit Model (DDIM), Song et al., ICLR 2021

* Progressive Distillation for Fast Sampling of Diffusion Models, Salimans et al., ICLR 2022

* Semantic Image Synthesis with Spatially-Adaptive Normalization (GauGAN), Park et al., CVPR 2019

* GAN Compression: Efficient Architectures for Interactive Conditional GANs, Li et al., CVPR 2020

* High-Resolution Image Synthesis with Latent Diffusion Models, Rombach et al., CVPR 2022



## Prerequisites



* Python3

* CPU, M1 GPU, or NVIDIA GPU + CUDA CuDNN

* [PyTorch](https://pytorch.org) >= 1.7. For M1 GPU support, please install [PyTorch](https://pytorch.org)>=2.0.



## Getting Started



### Installation



After installing [PyTorch](https://pytorch.org), you should be able to install SIGE with PyPI



```shell

pip install sige

```



or via GitHub:



```shell

pip install git+https://github.com/lmxyy/sige.git

```



or locally for development



```shell

git clone [email protected]:lmxyy/sige.git

cd sige

pip install -e .

```



For MPS backend, please set the environment variables:



```shell

export PYTORCH_ENABLE_MPS_FALLBACK=1

```



### Usage Example



See [example.py](https://github.com/lmxyy/sige/tree/main/example.py) for the minimal SIGE convolution example. Please first install SIGE with the above instructions and [torchprofile](https://github.com/zhijian-liu/torchprofile) with



```shell

pip install torchprofile

```



Then you can run it with



```shell

python example.py

```



We also have [![colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/lmxyy/sige/blob/main/example.ipynb) example.



### Benchmark



To reproduce the results of [DDPM](https://github.com/ermongroup/ddim) and [Progressive Distillation](https://github.com/google-research/google-research/tree/master/diffusion_distillation) or download the LSUN Church editing datasets, please follow the instructions in [diffusion/README.md](https://github.com/lmxyy/sige/tree/main/diffusion/README.md).



To reproduce the results of [GauGAN](https://github.com/NVlabs/SPADE) and [GAN Compression](https://github.com/mit-han-lab/gan-compression) or download the Cityscapes editing datasets, please follow the instructions in [gaugan/README.md](https://github.com/lmxyy/sige/tree/main/gaugan/README.md).



## Citation



If you use this code for your research, please cite our paper.



```bibtex

@inproceedings{li2022efficient,

  title={Efficient Spatially Sparse Inference for Conditional GANs and Diffusion Models},

  author={Li, Muyang and Lin, Ji and Meng, Chenlin and Ermon, Stefano and Han, Song and Zhu, Jun-Yan},

  booktitle={Advances in Neural Information Processing Systems (NeurIPS)},

  year={2022}

}

```



## Acknowledgments



Our code is developed based on [SDEdit](https://github.com/ermongroup/SDEdit), [ddim](https://github.com/ermongroup/ddim), [diffusion_distillation](https://github.com/google-research/google-research/tree/master/diffusion_distillation), [gan-compression](https://github.com/mit-han-lab/gan-compression), [dpm-solver](https://github.com/LuChengTHU/dpm-solver), and [stable-diffusion](https://github.com/CompVis/stable-diffusion). We refer to [sbnet](https://github.com/uber-research/sbnet) for the tiling-based sparse convolution algorithm implementation. Our work is also inspired by the gather/scatter implementations in [torchsparse](https://github.com/mit-han-lab/torchsparse).



We thank [torchprofile](https://github.com/zhijian-liu/torchprofile) for MACs measurement, [clean-fid](https://github.com/GaParmar/clean-fid) for FID computation and [drn](https://github.com/fyu/drn) for Cityscapes mIoU computation.



We thank Yaoyao Ding, Zihao Ye, Lianmin Zheng, Haotian Tang, and Ligeng Zhu for the helpful comments on the engine design. We also thank George Cazenavette, Kangle Deng, Ruihan Gao, Daohan Lu, Sheng-Yu Wang and Bingliang Zhang for their valuable feedback. The project is partly supported by NSF, MIT-IBM Watson AI Lab, Kwai Inc, and Sony Corporation.