CHI 2025

🍳GenPara

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

Jiin Choi1,2, Seung Won Lee1,2, Kyung Hoon Hyun1,2
1Design Informatics Lab, Interior Architecture Design, Hanyang University, Seoul, Republic of Korea
2Human-Centered AI Design Institute, Hanyang University, Seoul, Republic of Korea
ACM CHI Conference on Human Factors in Computing Systems (CHI ’25)
Paper arXiv Video Code

TL;DR

An interactive 3D editing system that turns text into shape parameters — and infers what part of the design space you actually care about.

Specifying design objectives and visualizing complex shapes through text alone is hard. GenPara leverages text-conditional shape parameters of part-aware 3D designs and lays the design space out in an Exploration Map and a Design Versioning Tree. Among the many shape parameters an LLM generates, the system uses Bayesian inference to surface outcomes within the user’s regions of interest. A user study (N = 16) showed GenPara improves comprehension and management of text-conditional shape parameters, streamlining exploration and concretization — boosting both efficiency and creativity in the 3D design process.

GenPara overview: an interactive 3D design editing system that leverages text-conditional shape parameters and infers users' regions of interest.
Overview. GenPara, an interactive 3D design editing system that leverages text-conditional shape parameters of 3D design and infers the users’ regions of interest.

Abstract

Bridging text and part-aware shape parameters for 3D design.

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 struggle 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 the design space within the Exploration Map and Design Versioning Tree. Additionally, among the various shape parameters generated by an 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 GenPara enhanced the comprehension and management of designers with text-conditional shape parameters, streamlining design exploration and concretization. This improvement boosted the efficiency and creativity of the 3D design process.

Video

See GenPara in action.

The problem

Challenges and research goals.

Text alone underspecifies 3D intent. GenPara is designed around the core challenges of editing part-aware designs through shape parameters.

Core challenges in 3D design editing with text-conditional shape parameters.
Core challenges.
Research goals derived from the core challenges.
Research goals.

The interface

Key tools.

Three tightly coupled capabilities let designers edit precisely, navigate the design space, and track how a design evolves.

tool 01

Text-conditional part-level editing

Precise shape modifications using shape parameters, conditioned on text at the level of individual parts.

tool 02

Visualization of the design space

An Exploration Map helps users navigate design variations effectively rather than guessing in the dark.

tool 03

Hierarchical visualization of designs

A Design Versioning Tree tracks and refines design evolution systematically across iterations.

The GenPara interface: Exploration Map and Design Versioning Tree for navigating the design space.
Key tools. The Exploration Map and Design Versioning Tree.

Generation & editing

LLM-based 3D generation and editing

A fine-tuned LLM produces text-conditional shape parameters; GenPara generates and edits part-aware 3D designs from them.

Fine-tuned LLM generating text-conditional shape parameters.
Fine-tuned generation.
GenPara 3D design generation and editing results.
Results.
GenPara editing an airplane 3D design through text-conditional shape parameters.
Part-aware editing example.

User study

Sixteen designers, GenPara vs. a prompt-only baseline.

A within-subjects study with 16 design-major participants (7 women, 9 men; M age = 24.6) from architecture, construction, interior, and lighting design — 14 had prior GenAI experience. Each explored and concretized a chair design for a given interior space using GenPara and a Baseline without its key features (Exploration Map, LLM design alternatives, Design Versioning Tree), followed by a 20-minute semi-structured interview.

RQ1 · Comprehension & visualization

GenPara helped designers grasp the relationship between text prompts and shape parameters. Its intuitive visualization tools significantly aided understanding during exploration and goal concretization.

Exploration (CSI): 8.50 vs. 5.06 Baseline (p<.001)
RQ2 · Prompting effort & specification

Part-based editing let designers iteratively refine a single design instead of picking among undefined results — expressing detailed intent with far fewer prompts, despite spending more time per session.

Prompt count: 8 vs. 2 · Effort (TLX): 2.94 vs. 4.25 (p<.05)
RQ3 · Creativity & efficiency

Designers reported lower frustration, higher transparency, and a significantly greater willingness to keep using GenPara — rating part-level designs more highly than prompt-only ones.

Enjoyment (CSI): 8.38 vs. 5.94 (p<.01)
Survey item GenPara Baseline Sig.
CSI · Exploration8.505.06***
CSI · Enjoyment8.385.94**
CSI · Results Worth Effort8.135.50**
CSI · Immersion7.946.25*
CSI · Expressiveness6.253.75**
NASA-TLX · Performance5.814.06***
NASA-TLX · Effort (lower is better)2.944.25*
NASA-TLX · Frustration (lower is better)1.753.06*

CSI on a 10-point scale, NASA-TLX on a 7-point scale. * p<.05   ** p<.01   *** p<.001.

Physical demand was slightly higher with GenPara (2.69 vs. 1.75) due to its detailed, part-specific editing — the one cost of finer control. The extra time spent specifying goals translated into greater immersion and more thorough concretization.

In their words

What the designers said.

Excerpts from the in-depth interviews, on understanding the design space, expressing what text can’t, and structuring the process.

“Through the Exploration Map, I could intuitively understand which design I was viewing and which area I was exploring. This contrasted with Baseline, which only presented results — it’s easy to facilitate the ideation of how to concretize the design.”
— P3
“The design alternatives provided by GenPara visually represented complex parts effectively, greatly aiding in understanding details that were difficult to explain in text.”
— P7
“After quickly refining the design I initially envisioned, I used the Exploration Map to explore alternatives. Watching the colors update, I realized I was focusing too much on one area and readjusted what I needed to focus on.”
— P15
“Visualizing the changes in parameters throughout my process let me retrospectively review the overall process and assist in decision-making for future steps — making modifications to parts while considering shape transitions, for a diverse view.”
— P16, on the Design Versioning Tree
“When generating an ‘armchair,’ the design was positioned within a similar cluster. But a ‘thin-framed chair’ spread across several clusters — making it intuitive to grasp the diversity and trends of shape parameters. This helped me quickly set my initial design goal.”
— P12
“With Baseline, I had to modify prompts multiple times to obtain the chair with an extended leg rest, and this process was quite cumbersome and time-consuming.”
— P10

Discussion · implications

Exploration & concretization with shape parameters

Comparing options in a visually structured design space leads to more informed decisions and helps designers concretize ideas that are hard to put into text.

A novel LLM interaction in 3D space

Designers edit detailed modifications by shape parameter — without deep 3D-modeling expertise — turning the LLM into a partner for creative dialogue rather than a one-shot generator.

Expanding GenPara’s impact

Participants saw GenPara replacing time-consuming sketching/modeling and supporting client communication — with interest in extending it to phones, automobiles, architecture, and lighting.

BibTeX

@inproceedings{choi2025genpara,
  title     = {GenPara: Enhancing the 3D Design Editing Process by Inferring
               Users' Regions of Interest with Text-Conditional Shape
               Parameters},
  author    = {Choi, Jiin and Lee, Seung Won and Hyun, Kyung Hoon},
  booktitle = {Proceedings of the 2025 CHI Conference on Human Factors
               in Computing Systems},
  pages     = {1--21},
  year      = {2025},
  publisher = {ACM},
}