Visitors in a small museum study gained more conceptual knowledge about Kaiping’s fortified towers when they could make AI images during a guided workshop. The image-making group also scored higher on a quiz a week later. Gen-Diaolou’s result offers a concrete example of how generating a scene can become part of learning about the place it depicts.[1]
The diaolou of Kaiping, in China’s Guangdong province, combine Chinese and Western architectural influences. Built as fortified homes, communal refuges and watchtowers, they carry the history of migration and money sent home. UNESCO’s World Heritage listing includes the towers’ village and agricultural surroundings; the setting helps explain the buildings.[2]
The tower stays; the surrounding story changes
Gen-Diaolou, described by Lei Han and colleagues in a CHI 2026 paper, begins with curated material about architectural types, decoration and preservation. An AI conversational guide takes the persona of Huang Bixiu, the builder of Ruishi Lou. Visitors then work through three themes: historical reconstruction, possible risks and future preservation.[1]
The creative module turns a short idea into a fuller prompt using a heritage knowledge base. Visitors can review that prompt before it goes to an image generator. Rules require the tower’s recognisable structure and its Kaiping setting to remain. Historical scenes must fit the 1930s; future scenarios have more freedom over time and surrounding activity. These are design constraints on generation, rather than a guarantee that every resulting image is historically correct.[1]
One museum participant imagined a tower standing in an empty desert. According to the researchers’ interview record, the system redirected the scene towards fields, houses and ancestral halls. Another participant proposed a freer mode. That exchange makes the educational choice visible: the software supplies historical context while also deciding which imaginative departures to permit.[1]
The machinery combined DeepSeek v3 for conversation and prompt development with FLUX.1 Kontext Pro for images. The pictures began from a base rendering of a diaolou. A scene of decay, flooding or future reuse was therefore a way to explore an idea; the study did not establish these images as engineering assessments or forecasts of damage.[1]
What the comparison actually measured
At the Jiangmen Wuyi Museum of Overseas Chinese, 26 adults were assigned to two groups of 13 using stratified block randomisation. Both groups used the same knowledge module, including its conversational guide. One then expressed the three themes through sketches and short descriptions; the other used the image-generation module. Each condition ran in a separate session, with researchers facilitating tasks and group discussions.[1]
Learning scores in the museum experiment
| Measure | Sketches and descriptions | AI image generation |
|---|---|---|
| Immediate factual gain, 10-item section | +2.77 items | +2.38 items |
| Immediate conceptual gain, 10-item section | +1.31 items | +2.15 items |
| One-week quiz, score out of 10 | 7.46 | 8.77 |
The researchers reported statistically significant differences for conceptual gains and the delayed quiz, but no significant difference in factual gains. The extra activity appears to have helped with understanding and subsequent quiz performance in this setting. It did not demonstrate that generating images adds more factual recall than studying the shared material and sketching.[1]
The scheduled introduction, tasks and post-workshop assessment totalled two hours. Participants were aged 20–43, and 23 of the 26 were women. This was a small, facilitated workshop with a one-week follow-up; it cannot establish what happens during an ordinary museum visit or whether participants later help preserve a building.[1]
A useful design question for museums
For a museum considering this approach, the promising unit is the whole learning activity: reliable material, a specific imaginative task, an inspectable prompt and a conversation about the resulting scene. Simply placing an image generator beside an exhibit would leave out much of what participants did here.[1]
The desert example suggests an especially useful question to ask of a prototype: when it redirects a visitor’s idea, does it explain the historical reason? A correction that teaches why the surrounding village matters can be part of the encounter with the object. The generated picture then gives the group something specific to discuss.[1]
Sources & further reading
- Gen-Diaolou: An Integrated AI-Assisted Interactive System for Diachronic Understanding and Preservation of the Kaiping Diaolou Lei Han and colleagues, CHI 2026 (arXiv) · 3 Feb 2026 · Accessed 19 Oct 2026
- Kaiping Diaolou and Villages UNESCO World Heritage Centre · Accessed 19 Oct 2026