Cornell researchers have developed an image-generation method designed to create more accurate scenes involving multiple people and their interactions.
The system builds scenes one person at a time and uses each person's pose to help generate the next part of the image, reducing the need for users to supply pose information.
Wenxuan Peng, a doctoral student in computer science, presented the work at the Association for Computing Machinery's SIGGRAPH 2026 conference in Los Angeles. Co-authors were Hadar Averbuch-Elor and Bharath Hariharan.
The researchers used the FLUX image generator as the foundation for their method and trained it with data from Who's Waldo, a large vision-language dataset created at Cornell and Cornell Tech.
The model combines pose detection with a multimodal large language model to arrange descriptions, poses and spatial regions for each person. It starts with the primary figure, then adds nearby people and background figures while carrying pose information forward at each stage.
The team also developed a benchmark called DrawWaldoWorlds to evaluate whether models correctly depict who is doing what to whom. In experiments and a study involving 20 people, participants preferred the Cornell method by an average of 2-to-1 over two versions of FLUX.
Make FL1 part of your day.
Local reporting. Useful updates. More ways to stay close to your community.
Go ad-free
Support independent local reporting and enjoy FL1 without display ads.
$4.99 / month
Become a member →Cancel anytime. Third-party media players may contain ads.Get the daily briefing
Start your morning with the headlines that matter across the Finger Lakes.
Find more FL1 on Google
Choose FingerLakes1.com as a preferred source to see more of our reporting in your Google news results.






