Important Note
If you are interested in this project, please contact me at wangshengwu01 [at] gmail [dot] com. We may collaborate on this project.
Many people struggle to capture photos that look natural, aesthetically pleasing, or express the intended mood. Amateur phtographers often cannot provide clear posing directions, and subjects frequently feel awkward or unsure how to stand or smile. Existing pose guide apps offer static catalogs of suggested poses, requiring manual browsing and imitation — they do not adapt to the actual model’s posture in real time.
We envision Pose Recommender, a system that guides users to pose themselves well that considers both the intrinsic user factors and the extrinsic environment factors in real time.
Challenges
There is no such thing as good or bad aesthetic quality. There are some general rules that can be applied to certain cases, but special cases make up the majority of the cases.
What is a good pose? A good pose captures the user's individual characteristics and the environment. To get a good pose, we need to understand the user's individual characteristics. We also need to know the purpose of the photo as different using cases may require different poses and decide how formal it should be. In addition, as a crucial component of a pose, the user's facial expression is also important. How can we read the mood from the user's facial expression and body movement? And vice versa, how can we use mood to guide the user to pose?
Related works
可颂 offers a feature called "灵感跟拍" that allows users to select an example photo and the app would detect the head of the model and guide the model to pose like the example photo. It is a good attempt to solve the problem, but it lacks the ability to understand the user's individually characteristic beauty and help them get the best pose that suits them and the environment.
Tools that may help
There are a couple of tools/models from TensorFlow.js that may help.
Published on July 17, 2025