GAD Editorial · Published · · English study · Chinese summary · Study dated 2026-10-04
A furniture model becomes more valuable when a designer can change it without rebuilding its logic. Stretching a static object is easy; deciding which parts should move together is harder. An interior study often needs that second capability when testing several related sizes in the same room.
ShapeAssembly explores a program-based answer. First submitted on 17 September 2020, the research introduces a language that describes shapes through cuboid proxies, attachments, hierarchy and symmetry. A learned generative model produces programs, while adjustable variables allow a program structure to describe related shapes. The paper investigates shape synthesis and fitting, not a production furniture specification system. ShapeAssembly paper
Define a family through its changes
GAD proposes a narrow pilot for an interior-design team: study a family of shelving or seating proxies used during spatial planning. Start by naming the changes that the family must support. These might include overall width, spacing between repeated parts or the relationship between a supporting element and the surface above it.
Write the invariants alongside the variables. Which parts must stay attached? Which relationships should remain symmetric? Which dimension must remain fixed when another changes? This turns an abstract desire for an editable model into a small, testable design brief.
Use a manually created reference family as the comparison. The purpose is not to assume that learning will outperform conventional parametric modeling. It is to discover whether generated programs introduce useful arrangements while keeping the required relationships understandable and controllable.
Review the edit, not only the original
A plausible initial object is the beginning of the test. Apply a sequence of representative parameter changes and inspect the result after each one. Include the ends of the intended range, where an apparently sensible relationship may produce collisions, unexpected gaps or awkward proportions.
Keep a record of which changes were achieved through exposed parameters and which required rewriting the program. That distinction reveals whether the family is genuinely reusable for the team's task. A beautiful variant that needs extensive repair may still be worthwhile, but it should not be counted as an effortless parameter adjustment.
Ask another designer to explain the attachment hierarchy and make a specified change without help from the original author. Readability is part of editability. If only one person understands how the output works, the office has acquired another specialist artifact rather than a dependable shared resource.
Keep the physical questions separate
Cuboid proxies communicate structure at a simplified geometric level. They do not establish material thickness, fastening details, fabrication tolerances or load capacity. Before an option moves towards specification, replace assumptions with verified product information or appropriate design development.
Name the representation clearly in presentation files so a rough spatial proxy does not become an apparently resolved custom-furniture proposal through repetition. Keep manufacturer objects and research-derived studies distinguishable in the library.
GAD has not executed ShapeAssembly or validated its outputs. The proposed pilot focuses on an enduring architectural question raised by the research: what knowledge survives a change? A useful generated family preserves the relationships the designer intended, exposes the ones it cannot preserve and leaves the next person with an object they can understand.