AI image generators can create convincing pictures of individual people, but scenes involving several people interacting remain difficult. Cornell University researchers have developed a method that improves these images by incorporating each person’s pose into the generation process and progressively constructing the scene.
The approach was developed by doctoral student Wenxuan Peng, Hadar Averbuch-Elor, and Bharath Hariharan. Their research, Composing People Together: Iterative Pose-Image Generation for Multi-Person Interaction Scenes, was presented at SIGGRAPH 2026 in Los Angeles.
Instead of attempting to generate everyone simultaneously, the system builds a scene one person at a time. It first predicts the pose of the primary person. That information becomes a condition for predicting the next person, with the process continuing until the scene is complete. This iterative method helps the model understand relationships and actions involving multiple people.
The researchers used FLUX, an image generator developed by Black Forest Labs, as the foundation of their system. They trained their model using Who’s Waldo, a large-scale vision-language dataset containing thousands of images showing interactions among multiple individuals.
Pose detection is combined with a multimodal large language model to generate aligned descriptions, poses, and spatial regions for each person. The system organizes individuals logically, starting with the primary actor before adding nearby and background figures.
To evaluate performance, the researchers created DrawWaldoWorlds, a benchmark that tests whether generated images contain the required people and correctly represent their actions and relationships. This allows researchers to assess whether a model understands who is performing each action and toward whom.
Experiments and a 20-person user study showed that the Cornell approach outperformed existing methods. Participants preferred its results by an average ratio of two to one over two FLUX versions.
Importantly, users do not need to supply poses themselves. By automatically incorporating pose information, the method could make generating complex and believable multi-person interactions considerably easier.