People pose 20k title image

What is People posing 20k?

It is a new free ML training dataset specific for the use of unsupervised or semi-supervised generative deep learning algorithms. It contains 20015 clean images (extracted from the background) of people posing in different poses and with different camera angles and distance to the camera. The images have been crawled from the web, some synthetically created.

Why a new dataset?

Well there are already some larger generative datasets out there for faces for example (CelebaHQ and FFHQ) but generating faces is easy. On the other hand we have of course the famous Imagenet dataset but that is just too cluttered with too many very different categories which makes the usage of complete unsupervised generative network close to unfeasible. So we thought we need something in between that is more challenging than faces but at the same time not too broad to keep unsupervised generative deep learning improving. Additionally 20000 images is also a more challenging in terms of # of images.

Where to get?

You can download the dataset from here


You are free to use the dataset for research and personal usage. For commercial usage you have to get first the written permission from Mut1ny (see email contact further below)


We have tried out a few SOTA deep learning generative architectures, you can see the results below. If you have your own example and want to share feel free to send your results to us.

Examples using Progressive Growing Gan
Examples using Progressive Growing Gan
Example using GLOW
Example using GLOW
Example using StyleGAN
Example using StyleGAN


If you want to support us in this project why not become a Patreon on our Patreon page


You can contact us at