🍳 GenPara : Enhancing 3D Design Editing with Text-Conditional Shape Parameters and Bayesian ROI Inference

[Arxiv]
Jiin Choi, Seung Won Lee, Kyung Hoon Hyun

📄 Paper & Appendix [url]

🎉 This paper got accepted to [CHI2025]!

• Main Paper: GenPara: Enhancing the 3D Design Editing Process by Inferring Users’ Regions of Interest with Text-Conditional Shape Parameters
• Appendix: Dataset format, fine-tuning structure, HD evaluation metrics

Introduction

🍳 GenPara is an interactive 3D design-editing system that fine-tunes an LLM using Gaussian-blob–based 3D shape parameters and infers each user's Region of Interest (ROI) via Bayesian inference.

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 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.

UPDATES

• Upload of extrinsic latent dataset format and examples. (Nov 2025)
• Upload of LLM fine-tuning dataset structure and instructions. (Nov 2025)
• Release of full README with setup & usage guide. (Nov 2025)
• Generation Code with fine-tuned model (Nov 2025)
• Release of GenPara backend (LLM inference + ROI model). (Coming soon)
• Release of GenPara frontend (Exploration Map & Versioning Tree). (Coming soon)
• Public demo

✨ Key Features

🔹 1. Text-Conditional Part-Level Editing

• Shape edits based on adjectives (e.g., open, curved, thin)
• LLM outputs modified extrinsic latents for 16 Gaussian blobs

🔹 2. Exploration Map (UMAP)

• Visualizes large-scale 3D design space (~58k chairs)
• Differentiates user-generated vs. LLM-generated vs. ROI-region samples

🔹 3. Bayesian ROI Inference

• Identifies user interest regions based on recent selections

🔹 4. Design Versioning Tree

• Hierarchical visualization of all variations and their parent–child relationships

---

🎯 Text-guided latent editing & Generation with Fine-tuned Model

GenPara supports fine-tuning using extrinsic latent vectors generated by SALAD.
Please use ready-to-use data and pre-trained SALAD and SPAGHETTI checkpoints [here].

1. Prepare Extrinsic Latent Files

Each model should contain:

• 16 Gaussian blobs
• Each blob: 16 parameters (Mu, eigenvectors, Pi, eigenvalues)

Example JSON files in:

dataset/GenPara_finetuningDataset_chairs.jsonl/

2. Create Fine-tuning Dataset (JSONL)

Each record must contain:

• System instruction
• User request describing parts + adjectives
• Assistant output providing adjusted extrinsic latents

code/genpara_system_instruction.txt/

3. Fine-tune Model

Use OpenAI fine-tuning API:

openai api fine_tuning.jobs.create -t dataset/GenPara_finetuningDataset_chairs.jsonl -m gpt-3.5-turbo-1106

4. Deploy Fine-tuned Model

Model returns full extrinsic latent vectors suitable for reconstruction.

code/genpara_text_guided_chair_edit.py

🗺️ Frontend (Exploration Map & Versioning Tree)

to be appear

---

📚 Citation

@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}

---

🙌 Acknowledgements

GenPara builds on:

• SPAGHETTI
• SALAD
• UMAP
• OpenAI GPT Fine-tuning