Home

Awesome

clean-fid for Evaluating Generative Models

<br> <p align="center"> <img src="https://raw.githubusercontent.com/GaParmar/clean-fid/main/docs/images/cleanfid_demo_folders.gif" /> </p>

Downloads Downloads

Project | Paper | Slides | Colab-FID | Colab-Resize | Leaderboard Tables <br> Quick start: Calculate FID | Calculate KID

[New] Computing the FID using CLIP features [Kynkäänniemi et al, 2022] is now supported. See here for more details.

The FID calculation involves many steps that can produce inconsistencies in the final metric. As shown below, different implementations use different low-level image quantization and resizing functions, the latter of which are often implemented incorrectly.

<p align="center"> <img src="https://raw.githubusercontent.com/GaParmar/clean-fid/main/docs/images/resize_circle.png" width="800" /> </p>

We provide an easy-to-use library to address the above issues and make the FID scores comparable across different methods, papers, and groups.

FID Steps


Corresponding Manuscript

On Aliased Resizing and Surprising Subtleties in GAN Evaluation <br> Gaurav Parmar, Richard Zhang, Jun-Yan Zhu<br> CVPR, 2022 <br> CMU and Adobe

If you find this repository useful for your research, please cite the following work.

@inproceedings{parmar2021cleanfid,
  title={On Aliased Resizing and Surprising Subtleties in GAN Evaluation},
  author={Parmar, Gaurav and Zhang, Richard and Zhu, Jun-Yan},
  booktitle={CVPR},
  year={2022}
}

<br>

Aliased Resizing Operations <br>

The definitions of resizing functions are mathematical and <em>should never be a function of the library being used</em>. Unfortunately, implementations differ across commonly-used libraries. They are often implemented incorrectly by popular libraries. Try out the different resizing implementations in the Google colab notebook here.

<img src="https://raw.githubusercontent.com/GaParmar/clean-fid/main/docs/images/resize_circle_extended.png" width="800" /> <br>

The inconsistencies among implementations can have a drastic effect of the evaluations metrics. The table below shows that FFHQ dataset images resized with bicubic implementation from other libraries (OpenCV, PyTorch, TensorFlow, OpenCV) have a large FID score (≥ 6) when compared to the same images resized with the correctly implemented PIL-bicubic filter. Other correctly implemented filters from PIL (Lanczos, bilinear, box) all result in relatively smaller FID score (≤ 0.75). Note that since TF 2.0, the new flag antialias (default: False) can produce results close to PIL. However, it was not used in the existing TF-FID repo and set as False by default.

<p align="center"><img src="https://raw.githubusercontent.com/GaParmar/clean-fid/main/docs/images/table_resize_sc.png" width="500" /></p>

JPEG Image Compression

Image compression can have a surprisingly large effect on FID. Images are perceptually indistinguishable from each other but have a large FID score. The FID scores under the images are calculated between all FFHQ images saved using the corresponding JPEG format and the PNG format.

<p align="center"> <img src="https://raw.githubusercontent.com/GaParmar/clean-fid/main/docs/images/jpeg_effects.png" width="800" /> </p>

Below, we study the effect of JPEG compression for StyleGAN2 models trained on the FFHQ dataset (left) and LSUN outdoor Church dataset (right). Note that LSUN dataset images were collected with JPEG compression (quality 75), whereas FFHQ images were collected as PNG. Interestingly, for LSUN dataset, the best FID score (3.48) is obtained when the generated images are compressed with JPEG quality 87.

<p align="center"> <img src="https://raw.githubusercontent.com/GaParmar/clean-fid/main/docs/images/jpeg_plots.png" width="800" /> </p>

Quick Start

Computing FID

Computing CLIP-FID

To use the CLIP features when computing the FID [Kynkäänniemi et al, 2022], specify the flag model_name="clip_vit_b_32"

Computing KID

The KID score can be computed using a similar interface as FID. The dataset statistics for KID are only precomputed for smaller datasets AFHQ, BreCaHAD, and MetFaces.


Supported Precomputed Datasets

We provide precompute statistics for the following commonly used configurations. Please contact us if you want to add statistics for your new datasets.

TaskDatasetResolutionReference Split# Reference Imagesmode
Image Generationcifar1032train50,000clean, legacy_tensorflow, legacy_pytorch
Image Generationcifar1032test10,000clean, legacy_tensorflow, legacy_pytorch
Image Generationffhq1024, 256trainval50,000clean, legacy_tensorflow, legacy_pytorch
Image Generationffhq1024, 256trainval70k70,000clean, legacy_tensorflow, legacy_pytorch
Image Generationlsun_church256train50,000clean, legacy_tensorflow, legacy_pytorch
Image Generationlsun_church256trainfull126,227clean
Image Generationlsun_horse256train50,000clean, legacy_tensorflow, legacy_pytorch
Image Generationlsun_horse256trainfull2,000,340clean
Image Generationlsun_cat256train50,000clean, legacy_tensorflow, legacy_pytorch
Image Generationlsun_cat256trainfull1,657,264clean, legacy_tensorflow, legacy_pytorch
Few Shot Generationafhq_cat512train5153clean, legacy_tensorflow, legacy_pytorch
Few Shot Generationafhq_dog512train4739clean, legacy_tensorflow, legacy_pytorch
Few Shot Generationafhq_wild512train4738clean, legacy_tensorflow, legacy_pytorch
Few Shot Generationbrecahad512train1944clean, legacy_tensorflow, legacy_pytorch
Few Shot Generationmetfaces1024train1336clean, legacy_tensorflow, legacy_pytorch
Image to Imagehorse2zebra256test140clean, legacy_tensorflow, legacy_pytorch
Image to Imagecat2dog256test500clean, legacy_tensorflow, legacy_pytorch

Using precomputed statistics In order to compute the FID score with the precomputed dataset statistics, use the corresponding options. For instance, to compute the clean-fid score on generated 256x256 FFHQ images use the command:

fid_score = fid.compute_fid(fdir1, dataset_name="ffhq", dataset_res=256,  mode="clean", dataset_split="trainval70k")

Create Custom Dataset Statistics


Backwards Compatibility

We provide two flags to reproduce the legacy FID score.


Building clean-fid locally from source

python setup.py bdist_wheel
pip install dist/*

CleanFID Leaderboard for common tasks

We compute the FID scores using the corresponding methods used in the original papers and using the Clean-FID proposed here. All values are computed using 10 evaluation runs. We provide an API to query the results shown in the tables below directly from the pip package.

If you would like to add new numbers and models to our leaderboard, feel free to contact us.

CIFAR-10 (few shot)

The test set is used as the reference distribution and compared to 10k generated images.

100% data (unconditional)

ModelLegacy-FID<br>(reported)Legacy-FID<br>(reproduced)Clean-FID
stylegan2 (+ada + tuning) [Karras et al, 2020]- †- †8.20 ± 0.10
stylegan2 (+ada) [Karras et al, 2020]- †- †9.26 ± 0.06
stylegan2 (diff-augment) [Zhao et al, 2020] [ckpt]9.899.90 ± 0.0910.85 ± 0.10
stylegan2 (mirror-flips) [Karras et al, 2020] [ckpt]11.0711.07 ± 0.1012.96 ± 0.07
stylegan2 (without-flips) [Karras et al, 2020]- †- †14.53 ± 0.13
AutoGAN (config A) [Gong et al, 2019]- †- †21.18 ± 0.12
AutoGAN (config B) [Gong et al, 2019]- †- †22.46 ± 0.15
AutoGAN (config C) [Gong et al, 2019]- †- †23.62 ± 0.30

† These methods use the training set as the reference distribution and compare to 50k generated images

20% data

ModelLegacy-FID<br>(reported)Legacy-FID<br>(reproduced)Clean-FID
stylegan2-diff-augment [Zhao et al, 2020] [ckpt]12.1512.12 ± 0.1514.18 ± 0.13
stylegan2-mirror-flips [Karras et al, 2020] [ckpt]23.0823.01 ± 0.1929.49 ± 0.17

10% data

ModelLegacy-FID<br>(reported)Legacy-FID<br>(reproduced)Clean-FID
stylegan2-diff-augment [Zhao et al, 2020] [ckpt]14.5014.53 ± 0.1216.98 ± 0.18
stylegan2-mirror-flips [Karras et al, 2020] [ckpt]36.0235.94 ± 0.1743.60 ± 0.17
<br>

CIFAR-100 (few shot)

The test set is used as the reference distribution and compared to 10k generated images.

100% data

ModelLegacy-FID<br>(reported)Legacy-FID<br>(reproduced)Clean-FID
stylegan2-mirror-flips [Karras et al, 2020] [ckpt]16.5416.44 ± 0.1918.44 ± 0.24
stylegan2-diff-augment [Zhao et al, 2020] [ckpt]15.2215.15 ± 0.1316.80 ± 0.13

20% data

ModelLegacy-FID<br>(reported)Legacy-FID<br>(reproduced)Clean-FID
stylegan2-mirror-flips [Karras et al, 2020] [ckpt]32.3032.26 ± 0.1934.88 ± 0.14
stylegan2-diff-augment [Zhao et al, 2020] [ckpt]16.6516.74 ± 0.1018.49 ± 0.08

10% data

ModelLegacy-FID<br>(reported)Legacy-FID<br>(reproduced)Clean-FID
stylegan2-mirror-flips [Karras et al, 2020] [ckpt]45.8745.97 ± 0.2046.77 ± 0.19
stylegan2-diff-augment [Zhao et al, 2020] [ckpt]20.7520.69 ± 0.1223.40 ± 0.09
<br>

FFHQ

all images @ 1024x1024<br> Values are computed using 50k generated images

ModelLegacy-FID<br>(reported)Legacy-FID<br>(reproduced)Clean-FIDReference Split
stylegan1 (config A) [Karras et al, 2020]4.44.39 ± 0.034.77 ± 0.03trainval
stylegan2 (config B) [Karras et al, 2020]4.394.43 ± 0.034.89 ± 0.03trainval
stylegan2 (config C) [Karras et al, 2020]4.384.40 ± 0.024.79 ± 0.02trainval
stylegan2 (config D) [Karras et al, 2020]4.344.34 ± 0.024.78 ± 0.03trainval
stylegan2 (config E) [Karras et al, 2020]3.313.33 ± 0.023.79 ± 0.02trainval
stylegan2 (config F) [Karras et al, 2020] [ckpt]2.842.83 +- 0.033.06 +- 0.02trainval
stylegan2 [Karras et al, 2020] [ckpt]N/A2.76 ± 0.032.98 ± 0.03trainval70k
<br>

140k - images @ 256x256 (entire training set with horizontal flips) The 70k images from trainval70k set is used as the reference images and compared to 50k generated images.

ModelLegacy-FID<br>(reported)Legacy-FID<br>(reproduced)Clean-FID
zCR [Zhao et al, 2020]3.45 ± 0.193.29 ± 0.013.40 ± 0.01
stylegan2 [Karras et al, 2020]3.66 ± 0.103.57 ± 0.033.73 ± 0.03
PA-GAN [Zhang and Khoreva et al, 2019]3.78 ± 0.063.67 ± 0.033.81 ± 0.03
stylegan2-ada [Karras et al, 2020]3.88 ± 0.133.84 ± 0.023.93 ± 0.02
Auxiliary rotation [Chen et al, 2019]4.16 ± 0.054.10 ± 0.024.29 ± 0.03
Adaptive Dropout [Karras et al, 2020]4.16 ± 0.054.09 ± 0.024.20 ± 0.02
Spectral Norm [Miyato et al, 2018]4.60 ± 0.194.43 ± 0.024.65 ± 0.02
WGAN-GP [Gulrajani et al, 2017]6.54 ± 0.376.19 ± 0.036.62 ± 0.03

† reported by [Karras et al, 2020] <br>

30k - images @ 256x256 (Few Shot Generation)<br> The 70k images from trainval70k set is used as the reference images and compared to 50k generated images.

ModelLegacy-FID<br>(reported)Legacy-FID<br>(reproduced)Clean-FID
stylegan2 [Karras et al, 2020] [ckpt]6.166.14 ± 0.0646.49 ± 0.068
DiffAugment-stylegan2 [Zhao et al, 2020] [ckpt]5.055.07 ± 0.0305.18 ± 0.032

10k - images @ 256x256 (Few Shot Generation)<br> The 70k images from trainval70k set is used as the reference images and compared to 50k generated images.

ModelLegacy-FID<br>(reported)Legacy-FID<br>(reproduced)Clean-FID
stylegan2 [Karras et al, 2020] [ckpt]14.7514.88 ± 0.07016.04 ± 0.078
DiffAugment-stylegan2 [Zhao et al, 2020] [ckpt]7.867.82 ± 0.0458.12 ± 0.044

5k - images @ 256x256 (Few Shot Generation)<br> The 70k images from trainval70k set is used as the reference images and compared to 50k generated images.

ModelLegacy-FID<br>(reported)Legacy-FID<br>(reproduced)Clean-FID
stylegan2 [Karras et al, 2020] [ckpt]26.6026.64 ± 0.08628.17 ± 0.090
DiffAugment-stylegan2 [Zhao et al, 2020] [ckpt]10.4510.45 ± 0.04710.99 ± 0.050

1k - images @ 256x256 (Few Shot Generation) <br> The 70k images from trainval70k set is used as the reference images and compared to 50k generated images.

ModelLegacy-FID<br>(reported)Legacy-FID<br>(reproduced)Clean-FID
stylegan2 [Karras et al, 2020] [ckpt]62.1662.14 ± 0.10864.17 ± 0.113
DiffAugment-stylegan2 [Zhao et al, 2020] [ckpt]25.6625.60 ± 0.07127.26 ± 0.077
<br>

LSUN Categories

100% data<br> The 50k images from train set is used as the reference images and compared to 50k generated images.

CategoryModelLegacy-FID<br>(reported)Legacy-FID<br>(reproduced)Clean-FID
Outdoor Churchesstylegan2 [Karras et al, 2020] [ckpt]3.863.87 ± 0.0294.08 ± 0.028
Horsesstylegan2 [Karras et al, 2020] [ckpt]3.433.41 ± 0.0213.62 ± 0.023
Catstylegan2 [Karras et al, 2020] [ckpt]6.937.02 ± 0.0397.47 ± 0.035
<br>

LSUN CAT - 30k images (Few Shot Generation)<br> All 1,657,264 images from trainfull split are used as the reference images and compared to 50k generated images.

ModelLegacy-FID<br>(reported)Legacy-FID<br>(reproduced)Clean-FID
stylegan2-mirror-flips [Karras et al, 2020] [ckpt]10.1210.15 ± 0.0410.87 ± 0.04
stylegan2-diff-augment [Zhao et al, 2020] [ckpt]9.689.70 ± 0.0710.25 ± 0.07

LSUN CAT - 10k images (Few Shot Generation)<br> All 1,657,264 images from trainfull split are used as the reference images and compared to 50k generated images.

ModelLegacy-FID<br>(reported)Legacy-FID<br>(reproduced)Clean-FID
stylegan2-mirror-flips [Karras et al, 2020] [ckpt]17.9317.98 ± 0.0918.71 ± 0.09
stylegan2-diff-augment [Zhao et al, 2020] [ckpt]12.0712.04 ± 0.0812.53 ± 0.08

LSUN CAT - 5k images (Few Shot Generation)<br> All 1,657,264 images from trainfull split are used as the reference images and compared to 50k generated images.

ModelLegacy-FID<br>(reported)Legacy-FID<br>(reproduced)Clean-FID
stylegan2-mirror-flips [Karras et al, 2020] [ckpt]34.6934.66 ± 0.1235.85 ± 0.12
stylegan2-diff-augment [Zhao et al, 2020] [ckpt]16.1116.11 ± 0.0916.79 ± 0.09

LSUN CAT - 1k images (Few Shot Generation)<br> All 1,657,264 images from trainfull split are used as the reference images and compared to 50k generated images.

ModelLegacy-FID<br>(reported)Legacy-FID<br>(reproduced)Clean-FID
stylegan2-mirror-flips [Karras et al, 2020] [ckpt]182.85182.80 ± 0.21185.86 ± 0.21
stylegan2-diff-augment [Zhao et al, 2020] [ckpt]42.2642.07 ± 0.1643.12 ± 0.16
<br>

AFHQ (Few Shot Generation)

AFHQ Dog<br> All 4739 images from train split are used as the reference images and compared to 50k generated images.

ModelLegacy-FID<br>(reported)Legacy-FID<br>(reproduced)Clean-FID
stylegan2 [Karras et al, 2020] [ckpt]19.3719.34 ± 0.0820.10 ± 0.08
stylegan2-ada [Karras et al, 2020] [ckpt]7.407.41 ± 0.027.61 ± 0.02

AFHQ Wild<br> All 4738 images from train split are used as the reference images and compared to 50k generated images.

ModelLegacy-FID<br>(reported)Legacy-FID<br>(reproduced)Clean-FID
stylegan2 [Karras et al, 2020] [ckpt]3.483.55 ± 0.033.66 ± 0.02
stylegan2-ada [Karras et al, 2020] [ckpt]3.053.01 ± 0.023.03 ± 0.02
<br>

BreCaHAD (Few Shot Generation)

All 1944 images from train split are used as the reference images and compared to 50k generated images.

ModelLegacy<br>FID<br>(reported)Legacy<br>FID<br>(reproduced)Clean-FIDLegacy<br>KID<br>(reported)<br>10^3Legacy<br>KID<br>(reproduced)<br>10^3Clean<br>KID<br>10^3
stylegan2 [Karras et al, 2020] [ckpt]97.7297.46 ± 0.1798.35 ± 0.1789.7689.90 ± 0.3192.51 ± 0.32
stylegan2-ada [Karras et al, 2020] [ckpt]15.7115.70 ± 0.0615.63 ± 0.062.882.93 ± 0.083.08 ± 0.08
<br>

MetFaces (Few Shot Generation)

All 1336 images from train split are used as the reference images and compared to 50k generated images.

ModelLegacy<br>FID<br>(reported)Legacy<br>FID<br>(reproduced)Clean-FIDLegacy<br>KID<br>(reported)<br>10^3Legacy<br>KID<br>(reproduced)<br>10^3Clean<br>KID<br>10^3
stylegan2 [Karras et al, 2020] [ckpt]57.2657.36 ± 0.1065.74 ± 0.1135.6635.69 ± 0.1640.90 ± 0.14
stylegan2-ada [Karras et al, 2020] [ckpt]18.2218.18 ± 0.0319.60 ± 0.032.412.38 ± 0.052.86 ± 0.04
<br>

Horse2Zebra (Image to Image Translation)

All 140 images from test split are used as the reference images and compared to 120 translated images.

ModelLegacy-FID<br>(reported)Legacy-FID<br>(reproduced)Clean-FID
CUT [Park et al, 2020]45.545.5143.71
Distance [Benaim and Wolf et al, 2017] reported by [Park et al, 2020]72.071.9771.01
FastCUT [Park et al, 2020]73.473.3872.53
CycleGAN [Zhu et al, 2017] reported by [Park et al, 2020]77.277.2075.17
SelfDistance [Benaim and Wolf et al, 2017] reported by [Park et al, 2020]80.880.7879.28
GCGAN [Fu et al, 2019] reported by [Park et al, 2020]86.785.8683.65
MUNIT [Huang et al, 2018] reported by [Park et al, 2020]133.8- †120.48
DRIT [Lee et al, 2017] reported by [Park et al, 2020]140.0- †99.56

† The translated images for these methods were intitially compared by [Park et al, 2020] using .jpeg compression. We retrain these two methods using the same protocal and generate the images as .png for a fair comparision.

<br>

Cat2Dog (Image to Image Translation)

All 500 images from test split are used as the reference images and compared to 500 translated images.

ModelLegacy-FID<br>(reported)Legacy-FID<br>(reproduced)Clean-FID
CUT [Park et al, 2020]76.276.2177.58
FastCUT [Park et al, 2020]94.093.9595.37
GCGAN [Fu et al, 2019] reported by [Park et al, 2020]96.696.6196.49
MUNIT [Huang et al, 2018] reported by [Park et al, 2020]104.4- †123.73
DRIT [Lee et al, 2017] reported by [Park et al, 2020]123.4- †127.21
SelfDistance [Benaim and Wolf et al, 2017] reported by [Park et al, 2020]144.4144.42147.23
Distance [Benaim and Wolf et al, 2017] reported by [Park et al, 2020]155.3155.34158.39

† The translated images for these methods were intitially compared by [Park et al, 2020] using .jpeg compression. We retrain these two methods using the same protocal and generate the images as .png for a fair comparision.


Related Projects

torch-fidelity: High-fidelity performance metrics for generative models in PyTorch. <br> TTUR: Two time-scale update rule for training GANs. <br> LPIPS: Perceptual Similarity Metric and Dataset. <br>


Licenses

All material in this repository is made available under the MIT License.

inception_pytorch.py is derived from the PyTorch implementation of FID provided by Maximilian Seitzer. These files were originally shared under the Apache 2.0 License.

inception-2015-12-05.pt is a torchscript model of the pre-trained Inception-v3 network by Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. The network was originally shared under Apache 2.0 license on the TensorFlow Models repository. The torchscript wrapper is provided by Tero Karras and Miika Aittala and Janne Hellsten and Samuli Laine and Jaakko Lehtinen and Timo Aila which is released under the Nvidia Source Code License.