Treellis Lets Designers Fix One Part Without Rewriting the Whole 3D Model

A prompt-generated model can look finished while still being frustrating to change. Ask for a slightly wider handle and a prompt-only generator may redraw the handle, the body and the proportions you had already accepted. Treellis tackles that revision problem by making the model’s parts and parameters visible, so a person can target the piece that needs work and keep the rest stable. ^[1]

From a mesh you receive to a structure you can change

The Treellis research team describes a system that generates a 3D object together with its assembly structure and object-specific parameters. The interface coordinates a prompt area, a 3D viewport, a structure tree and a parameter panel. Selecting a part in the tree or directly in the viewport highlights the same component, then lets the user adjust parameters such as position, rotation, scale, size or color. An AI edit can also be scoped to the selected subtree. ^[1]

That design changes the interaction contract. The model is not just asked to return another complete object; the user can point to a component and say what should change while leaving accepted components alone. This can reduce the “regenerate and inspect” loop that makes prompt-only tools unpredictable. It also makes the internal decomposition a visible object of collaboration: the person can see which part is being edited and how it relates to the whole. ^[1]

The technique is useful only when the assembly is meaningful. A model split into arbitrary fragments would expose controls without giving the designer a useful way to reason about the object. The researchers’ approach makes editable structure part of the output, so the quality of that structure becomes as important as the visual quality of the initial render.

What the reported evaluation does and does not say

The project page reports a controlled between-subjects study with 24 participants and a prompt-only baseline. Across 108 task attempts, Treellis participants completed 106 independently, compared with 92 in the baseline; the reported System Usability Scale scores were 87.7 and 74.0. A separate multi-session study involved five people making objects such as a toy car, bicycle, courtyard and heat exchanger. The work received a Best Paper Award at ACM UIST 2026. ^{[1] [2]}

These are concrete evaluations of an interface and a set of design tasks. They do not show that every generated object will be editable, that the tool produces dimensionally accurate parts, or that output can go directly into manufacturing. The project team itself frames possible clinical applications such as orthotics and splints as future directions. Those use cases require professional assessment, material constraints, fit testing and safety validation beyond the reported studies. ^[2]

The project is open source and the team provides a browser demo. That makes it possible to inspect the interaction model directly, but a demo’s availability is not the same as a production workflow for CAD, simulation or fabrication. ^{[1] [4]}

Why local editing matters beyond 3D

The useful idea is broader than a particular model generator: make the parts of an AI output addressable, then let a person edit the smallest relevant unit. In text, that may be one paragraph. In a spreadsheet, it may be a column formula. In a 3D scene, it can be a component with explicit geometry parameters. The smaller the edit boundary, the easier it is to preserve the work that is already correct and to review what changed.

This structure also creates a better place for accountability. A designer can inspect a selected component before asking for an AI revision and compare it afterward. That does not guarantee the model’s reasoning is transparent, but it gives the human a concrete object to verify. A good co-creation tool should make the target and the result visible rather than asking someone to trust that a broad prompt changed only the intended part.

For a team evaluating AI-assisted design, test a realistic repair task, not only first-generation quality. Ask a designer to adjust one component while preserving three others; record how many unrelated regions change, how often manual correction is needed, and whether the model stays usable for a second session. The benefit appears when revision is controlled, not when the first render looks impressive.

Treellis is a research prototype, but its core design principle is practical: preserve the parts a human has already approved, expose the structure needed for a precise edit, and keep the person responsible for deciding whether the new result is fit for its purpose. ^[1]

See also our coverage of AI-assisted image editing and creative workflows and Google’s chemistry agent workflow, where human review remains part of a more specialized process.