Related Open-Source Bases

gaussian-splatting

Foundational 3D Gaussian Splatting framework used by ForgeDreamer.

GraphDeco INRIA
Stable-Dreamfusion

Text-to-3D optimization recipe and training strategy inspiration.

Open Source Project
LoRA

Low-rank adaptation paradigm used in our multi-expert LoRA design.

Open Source Project

[CVPR 2026 Findings] ForgeDreamer: Industrial Text-to-3D Generation with Multi-Expert LoRA and Cross-View Hypergraph

Junhao Cai1,*, Deyu Zeng1,*, Junhao Pang1, Lini Li1, Zongze Wu1, Xiaopin Zhong1,†
1 Shenzhen University, Shenzhen, Guangdong 518060, China
2 Guangzhou Maritime University, Guangzhou, Guangdong 510725, China
CVPR 2026 Findings Industrial Text-to-3D Generation
Paper Code arXiv

ForgeDreamer integrates Multi-Expert LoRA distillation and Cross-View Hypergraph enhancement to improve geometry consistency, topology fidelity, and industrial detail quality in text-to-3D generation.

Abstract

Industrial text-to-3D generation is challenging due to strict geometric constraints, fine-grained topology requirements, and strong demands on structural consistency. We present ForgeDreamer, a framework tailored for high-fidelity industrial asset generation that combines Multi-Expert LoRA distillation with Cross-View Hypergraph enhancement. Multi-Expert LoRA improves domain-specific feature adaptation across diverse part categories, while the Cross-View Hypergraph explicitly models inter-view dependencies to preserve global coherence and local detail simultaneously. Extensive experiments on representative industrial components demonstrate that ForgeDreamer consistently improves geometry quality, topology faithfulness, and visual realism over strong baselines, providing a practical solution for industrial design prototyping and digital-twin content creation.

Industrial Cases (All GIFs)

Paper, Pipeline & Supplement

Pipeline Image · Paper PDF · Supplement PDF

ForgeDreamer pipeline

BibTeX

@article{cai2026forgedreamer,
  title={ForgeDreamer: Industrial Text-to-3D Generation with Multi-Expert LoRA and Cross-View Hypergraph},
  author={Cai, Junhao and Zeng, Deyu and Pang, Junhao and Li, Lini and Wu, Zongze and Zhong, Xiaopin},
  journal={arXiv preprint arXiv:2603.09266},
  year={2026},
  url={https://arxiv.org/abs/2603.09266}
}