Modern neural networks can turn a few words into a finished 3D object. But behind such a result there is often an unclear process: the user gets an item but does not understand how it was made. Researchers from the University of Sydney, together with colleagues from the University of Cambridge and Singapore Management University, have developed the Treellis tool, which makes collaboration with AI more transparent and educational.
The program does not work like conventional text-prompt generators. Instead of immediately producing the final design, Treellis offers to assemble the object step by step - manually or with the help of AI. The interface shows individual parts of the model and their arrangement in space. A person can select a specific element, change its size or position and immediately see the result. Mistakes and undoing actions are allowed - the researchers consider this an important part of learning.
"Treellis does not just do something for you. It shows how it is done," the university's press service quotes the lead author of the work, doctoral student Yixiang Chen, as saying.
The tool is distributed as open-source software: anyone with a browser can use it, and developers can integrate it into any AI platform via an application programming interface, or API - a set of rules by which programs communicate with each other.
To test whether Treellis helps people learn, a team led by Dr Anusha Withana conducted a trial with two groups. All participants had to create three 3D objects of increasing complexity: a cup, a table and a computer monitor - with specified requirements. One group worked with a conventional tool based on text prompts, the other with Treellis.
Observations showed that those who used Treellis actively explored how the model is assembled and tried small reversible changes to get closer to the desired result. This helped them intuitively grasp complex 3D design concepts. Participants in the group with the conventional generator often did not understand how the object was structured, and attempts to fix one part of the model created problems elsewhere.
According to Withana, the tool helps overcome two typical difficulties when working with AI: the difficulty of understanding how the system is structured, and the inability to perform the required task with its help. Interactivity, he believes, narrows both gaps and does not allow the work to be fully delegated to the algorithm.
The work was presented at the Association for Computing Machinery's Symposium on User Interface Software and Technology (UIST 26) in Michigan and received a best paper award. As noted by the chair of the award committee, Professor Anne Roudaut of the University of Bristol, the award recognizes originality, scientific rigor and the potential for long-term influence on the field.
In the future, the authors hope to apply the open interface in clinical practice. For example, doctors could use it to individually design foot and lower-leg orthoses for children with cerebral palsy, wrist splints for patients with arthritis or ergonomic handles for rehabilitation equipment. For now this is a laboratory development and plans, not a finished commercial product.
"Treellis breaks the pattern in which AI remains a black box. Using the example of 3D design, it allows users to understand the internal structure of the result, interact with it and change it," Withana concluded.