GenPara: Enhancing the 3D Design Editing Process by Inferring Users' Regions of Interest with Text-Conditional Shape Parameters

TL;DR

GenPara addresses the challenge of articulating and exploring complex 3D design goals by grounding text prompts in text-conditional shape parameters for part-aware 3D models. It combines a fine-tuned LLM, part-aware 3D GenAI (SPAGHETTI/SALAD), an Exploration Map for design-space navigation, and a Design Versioning Tree to capture design evolution, with Bayesian inference to infer the user’s ROI. A user study (N=16) shows GenPara enhances comprehension, reduces prompt-generation effort, and increases perceived creativity and efficiency compared to a Baseline gallery approach. The work demonstrates that structured visualization and ROI-aware generation can streamline 3D design exploration and concretization, enabling more targeted, iterative interactions between designers and GenAI. These findings suggest practical impact for early-stage design workflows and motivate extensions to other 3D domains and collaborative scenarios.

Abstract

In 3D design, specifying design objectives and visualizing complex shapes through text alone proves to be a significant challenge. Although advancements in 3D GenAI have significantly enhanced part assembly and the creation of high-quality 3D designs, many systems still to dynamically generate and edit design elements based on the shape parameters. To bridge this gap, we propose GenPara, an interactive 3D design editing system that leverages text-conditional shape parameters of part-aware 3D designs and visualizes design space within the Exploration Map and Design Versioning Tree. Additionally, among the various shape parameters generated by LLM, the system extracts and provides design outcomes within the user's regions of interest based on Bayesian inference. A user study N = 16 revealed that \textit{GenPara} enhanced the comprehension and management of designers with text-conditional shape parameters, streamlining design exploration and concretization. This improvement boosted efficiency and creativity of the 3D design process.

Figures (17)

• Figure 1: Overview of GenPara: This figure demonstrates how designers utilize GenPara to navigate and refine their design process. Designers generate initial chair concepts by providing text prompts (a1), supported by aligned and diversified adjectives (a2), in the Input Box (a). They explore potential designs on the (b) Exploration Map, represented as diverse dots. Users can select specific parts of a chair using 3D Gaussian blobs, as shown in (c1) 3D Gaussian Blob and Mesh, and switch between mesh and blob modes via (c2). The user uses (c3) (LLM Generation Button) to create and display new design alternatives in the Design Versioning Tree on the right (d), allowing users to review and evaluate the adjustments made to the chair design. This integrated approach supports a systematic exploration from conceptualization to detailed specification.
• Figure 2: Input Box: Users can generate chair designs via text by Input Box. The system generates adjective suggestions based on the prompt history of the user, such as “ergonomic,” “organic,” and “angular.” These suggestions enhance creative exploration by enabling users to effectively refine their design prompts, fostering the creation of customized 3D chair models that meet specific aesthetic and functional requirements.
• Figure 3: Exploration Map for the design space: This map visualizes chair designs, aiding efficient navigation through the design space. Designs generated by user prompts are marked in mint-quadrangle, whereas those from the LLM appear in yellow-triangle, enhancing user choice and exploration. As the user generates designs, the map indicates the ROI of each user in shades of pink, purple, and gray. To improve accessibility and ensure clear differentiation, the map in this paper is presented with shapes rather than solely relying on color, differing slightly from the UI used in the user study.
• Figure 4: Generation Card: Users can directly interact with the 3D chair model to specify modifications. (a): 3D Gaussian blob view of the chair allows users to interactively select parts they wish to edit. Users can rotate and examine these parts closely, enhancing understanding and precision in design customization; (b): 3D mesh view of the chair also shows the selected parts as it allows users to check parts they wish to edit. Users can toggle between 3D Gaussian blob mode for an abstract representation or mesh mode for a detailed view. Selected parts are highlighted in both modes to ensure clarity and continuity in the selection process. (c) LLM Generation Button: After making their selections, users press to generate the modified design alternatives. These are subsequently displayed in the Design Versioning Tree
• Figure 5: Design Versioning Tree: (a) 3D Gaussian mode, (b) Mesh mode; This figure displays a 3D node–link diagram visualizing tree-structured data of 3D design models generated by LLMs. Nodes represent individual design models while edges depict parent/child relationships organized along the y-axis to intuitively showcase the hierarchy of modified designs by users, facilitating an understanding of their focus in the design process.