New AI Model Developed by Stanford, UC Berkeley, and Adobe Research to Realistically Insert Specific Humans into Various Scenes

Have you ever wanted to insert yourself into your favorite movie scenes or holiday photographs? Generative models have made it possible for researchers to create computational tools capable of generating images or texts based on training data. Stanford University, UC Berkeley, and Adobe Research have now introduced a novel model that can seamlessly insert specific humans into different scenes with impressive realism. In this article, we will delve into the research of this advanced model, which holds significant potential for future research in the creative industries and robotics.

The Self-Supervised Training Approach:

The researchers employed a self-supervised training approach to teach the generative model how to realistically insert individuals into various scenes. The model was trained on videos featuring humans moving within different surroundings, and the researchers pursued the human’s potential poses through the training process. The model then put the human into another frame while still maintaining the same realistic poses.

Superior Performance:

The generative model outperformed non-generative models in terms of affordance perception, allowing the model to identify potential opportunities for interaction within an environment. This innovative model holds significant potential for the creative industries, supporting artists and media creators in enhancing their image editing functionalities and creating realistic images and videos.

Future Exploration:

The researchers are planning to incorporate greater controllability into the generated poses and expand the approach beyond just humans to encompass all objects. The addition of the model into photo editing smartphone applications would enable smartphone users to layer images with augmented human beings rapidly.


The researchers have introduced a new model that allows for the realistic insertion of humans into scenes. Leveraging generative models and self-supervised training, the model demonstrated significant capabilities and holds potential for various applications in the creative industries and robotics research. Future research will focus on refining and expanding the capabilities of the model. Do you want to learn more about this research? Check out their research paper and join our ML community and email newsletter to receive the latest AI research news and learn more about cool AI projects.

Leave a comment

Your email address will not be published. Required fields are marked *