Shisei is a free CC0 pose library for AI image tools

A new reference library for AI image tools opened on GitHub this week. It is called Shisei and the author is wquguru. The repository held 514 stars on 14 November. The library holds 280 mannequin poses. Each pose is drawn from five camera angles: front, three-quarter, side, above and below.

Every image sits in the public domain under CC0 1.0. The code is Apache-2.0. The author named it after a Japanese word for a figure's posture. The library is split into four series: full body in standard poses, upper body, full body with sexy framing, and duos. The site reads in four languages: Japanese, Chinese, English and Korean.

What the library holds

Each pose comes in four kinds of file. There is a black-and-white line drawing. There is an AI photo render of the same pose. 89 sets are in the current release. There is a depth map and a normal map. The maps are made for ControlNet. ControlNet is a way of giving an AI image tool a hint about the shape of the subject. The line art is the file an artist would hand to Nano Banana, ChatGPT, Midjourney or ComfyUI as a reference.

The files come in three web sizes: 300 pixels for a thumbnail, 600 by 800 for a preview, and 900 by 1200 for a full view. The full originals run at 1344 by 1792. The AI photo renders are larger, at 1120 by 2240. The whole archive holds 1,301 line art images and 944 map variants. A thumbnail is about 7 KB. A full file is about 100 KB.

What someone can do with it today

The live gallery is at shisei.indaloop.app. An artist can pick a pose and an angle. They can pull the line art or the depth map straight into the tool they already use. The same character can then appear in one pose across all five angles. The reference is fixed. Users can also download the whole set as a ZIP from GitHub Releases. The current release is v0.4.1, so the pose list can still change.

How it compares to other pose sources

A free pose library is not new. OpenPose skeletons have been shared for years. Shisei does one thing OpenPose does not. It holds five matched angles of the same pose. Every image is in the public domain. There is no attribution rule. A team can use any image in a released product and record no line about where the pose came from.

The library will not help a team that needs real people. The renders are mannequins. The author labels the project an early release. The pose set can keep growing. Old files can keep moving. A team using it in production would pin a release, not the main branch.

Source

Primary source: wquguru/shisei on GitHub. Live gallery: shisei.indaloop.app.