ForgeDreamer: Industrial Text-to-3D Generation with Multi-Expert LoRA and Cross-View Hypergraph

Junhao Cai; Deyu Zeng; Junhao Pang; Lini Li; Zongze Wu; Xiaopin Zhong

cs.CVMar 10, 2026

ForgeDreamer: Industrial Text-to-3D Generation with Multi-Expert LoRA and Cross-View Hypergraph

Authors: Junhao Cai, Deyu Zeng, Junhao Pang, Lini Li, Zongze Wu, Xiaopin Zhong

Organizations: Shenzhen University, Shenzhen, Guangdong 518060, China · Guangzhou Maritime University, Guangzhou, Guangdong 510725, China

Abstract

Current text-to-3D generation methods excel in natural scenes but struggle with industrial applications due to two critical limitations: domain adaptation challenges where conventional LoRA fusion causes knowledge interference across categories, and geometric reasoning deficiencies where pairwise consistency constraints fail to capture higher-order structural dependencies essential for precision manufacturing. We propose a novel framework named ForgeDreamer addressing both challenges through two key innovations. First, we introduce a Multi-Expert LoRA Ensemble mechanism that consolidates multiple category-specific LoRA models into a unified representation, achieving superior cross-category generalization while eliminating knowledge interference. Second, building on enhanced semantic understanding, we develop a Cross-View Hypergraph Geometric Enhancement approach that captures structural dependencies spanning multiple viewpoints simultaneously. These components work synergistically improved semantic understanding, enables more effective geometric reasoning, while hypergraph modeling ensures manufacturing-level consistency. Extensive experiments on a custom industrial dataset demonstrate superior semantic generalization and enhanced geometric fidelity compared to state-of-the-art approaches. Code is available at https://github.com/Junhaocai27/ForgeDreamer

Figures & tables

Figure 1: Overall Framework Architecture: Multi-Expert LoRA Ensemble Framework and Cross-View Hypergraph Enhancement

Figure 2: Overview of custom-built multi-view industrial dataset and 3D generation results. The bottom-right image presents the corresponding 3D generation result produced by our method.

Figure 3: Architecture of our industrial Text-to-3D generation framework - ForgeDreamer. Top: Cross-view Hypergraph Enhanced 3D Gaussian Generation pipeline with Cross-View Geometric Correlation Module (CVGCM). Bottom: Multi-Expert LoRA Ensemble Framework for cross-category industrial knowledge integration.

Figure 4: Qualitative Comparison with State-of-the-Art Methods. Visual results demonstrate the superior performance of our approach. See Appendix for remaining categories.

Method	Time †	Average ∗
ProlificDreamer [ 42 ] (w/o LoRA)	$\sim$ 10 hours	25.13
DreamFusion [ 33 ] (w/o LoRA)	6 hours	41.91
DreamFusion [ 33 ] (w/ LoRA)	6 hours	44.83
RichDreamer [ 34 ] (w/o LoRA)	120 minutes	28.27
LucidDreamer [ 20 ] (w/o LoRA)	110 minutes	47.10
LucidDreamer [ 20 ] (w/ LoRA)	110 minutes	46.75

Table 1: Condensed comparison of average scores (10 categories). † Total processing time; ∗ Average T3Bench [ 11 ] quality score. Best and second-best scores are bolded and underlined , respectively. See Appendix for full results.

Figure 5: LLM-based qualitative ranking across ten object prompts, showing that our method achieves the highest overall fidelity and consistency. Full evaluation results are included in the Appendix .

Figure 6: Multi-Expert LoRA Integration: Addition Fusion vs. Distillation Fusion

Method	Two LoRAs	Four LoRAs	Six LoRAs
Addition	0.938	0.814	0.633
Distillation	0.965	0.949	0.952

Table 2: Comparison of Average Concept Preservation Scores evaluated by CLIP-ViT-L/14. The full detailed table is presented in the Appendix .

Figure 7: Component-wise Ablation Analysis: LoRA (Distillation Fusion) vs. MVHG Loss vs. Combined Approach

Figure 8: Our data capture apparatus (left) and collected dataset samples (right). The setup utilizes controlled lighting and fixed mounts to acquire consistent front and top-down view images for each industrial component.

Panel of Figure 9 — Figure 9: Qualitative results on two public industrial datasets. (a) Real-IAD: our method accurately detects and localizes defects. (b) MVTec 3D-AD: our framework achieves improved defect completeness and surface consistency.

Figure 10: Qualitative Comparison with State-of-the-Art Methods. Visual results demonstrate the superior performance of our approach (remaining categories).

Individual Concepts
Method	Two	Four	Six
(CLIP-ViT-L/14)	LoRAs	LoRAs	LoRAs
Emb1 (Addition)	0.899	0.793	0.658
Emb1 (Distillation)	0.927	0.934	0.904
Emb2 (Addition)	0.886	0.779	0.653
Emb2 (Distillation)	0.915	0.963	0.963

Table 3: Concept Preservation Scores by Fusion Method and LoRA Configuration

Panel of Figure 11 — Figure 11: Analysis of Six LoRAs Configuration: (a) Cosine similarity comparison between addition fusion and ablation fusion methods. (b) PCA visualization showing the representational space distribution and relative distances to the original LoRA.

Method	G. LED	Nut	R. LED	Resistor	Screw	Bearing	Capacitor	Hex Stud	Nail	Gasket	Average
ProlificDreamer [ 42 ] (w/o LoRA)	7	7	7	7	7	7	7	7	7	6	6.9
RichDreamer [ 34 ] (w/o LoRA)	6	6	6	6	6	4	6	6	4	3	5.3
DreamFusion [ 33 ] (w/ LoRA)	3	5	2	2	2	2	4	2	1	1	2.4
DreamFusion [ 33 ] (w/o LoRA)	5	4	5	5	5	1	1	5	5	4	4.0
LucidDreamer [ 20 ] (w/LoRA)	2	3	3	3	4	6	5	4	3	5	3.8
LucidDreamer [ 20 ] (w/o LoRA)	4	2	4	4	3	5	2	3	6	7	4.0

Table 4: Quantitative user study ranking across ten object categories. Lower ranks are better. Best (rank 1) and second-best (rank 2) results in each column are highlighted in bold and underlined , respectively.

Method	G. LED	Nut	R. LED	Resistor	Screw	Bearing	Capacitor	Hex Stud	Nail	Gasket	Average
ProlificDreamer [ 42 ] (w/o LoRA)	27.19	22.34	24.49	26.60	34.85	31.64	20.48	7.85	26.39	29.45	25.13
RichDreamer [ 34 ] (w/o LoRA)	17.21	42.26	15.58	42.31	38.33	21.24	17.83	24.38	39.90	23.70	28.27
DreamFusion [ 33 ] (w/o LoRA)	32.62	69.72	29.92	28.49	46.59	28.38	35.45	36.66	52.27	59.02	41.91
DreamFusion [ 33 ] (w/ LoRA)	48.27	63.31	31.98	38.40	56.29	22.76	21.12	50.61	62.27	53.27	44.83
LucidDreamer [ 20 ] (w/o LoRA)	44.46	71.37	31.24	53.23	53.97	36.86	45.51	58.93	39.66	35.72	47.10
LucidDreamer [ 20 ] (w/ LoRA)	48.45	64.62	53.98	30.16	57.31	20.60	22.45	54.23	57.39	58.30	46.75

Table 5: Quantitative comparison across ten object categories on T3Bench quality scores. Best and second-best results in each column are highlighted in bold and underlined , respectively.

Figure 14: Qualitative comparison with state-of-the-art methods on natural scenes. Visual results demonstrate the superior performance of our approach.

Panel of Figure 15 — Figure 15: LLM-based qualitative evaluation (Part 1-3).

Panel of Figure 16 — Figure 16: LLM-based qualitative evaluation (Part 4-6).

Panel of Figure 17 — Figure 17: LLM-based qualitative evaluation (Part 7-9).

Panel of Figure 18 — Figure 18: LLM-based qualitative evaluation (Part 10-11).

Text-to-3D Generation

Text-to-3D generation creates 3D objects and scenes from language while balancing semantic alignment, geometry, appearance, and spatial coherence.

See all 64 papers

DreamPartGen: Semantically Grounded Part-Level 3D Generation via Collaborative Latent Denoising
Mar 19, 2026Tianjiao Yu, Xinzhuo Li, Muntasir Wahed +4Part-Level 3D Generation 3D Shape Generation
University of Illinois Urbana-Champaign
ReConText3D: Replay-based Continual Text-to-3D Generation
Apr 15, 2026Muhammad Ahmed Ullah Khan, Muhammad Haris Bin Amir, Didier Stricker +1Text-to-3D Generation 3D Asset Generation
DFKI · RPTU Kaiserslautern-Landau
Arko-T: A Foundation Model for Text-to-Structured 3D Generation
Jun 29, 2026Liang Wang, Zhaoyang Xi, Zekai Xiang +53D Foundation Models Parametric CAD Modeling
Spatial Design Intelligence Lab, BitInf Ltd., Shanghai 200003, China · School of Computer Science, Wuhan University, Wuhan 430000, Hubei, China · College of Computer and Information Engineering, Nanjing Tech University, Nanjing 211800, Jiangsu, China

Mar 19, 2026J/K move · Enter open · S save

DreamPartGen: Semantically Grounded Part-Level 3D Generation via Collaborative Latent Denoising

Tianjiao Yu, Xinzhuo Li, Muntasir Wahed, Jerry Xiong+3

University of Illinois Urbana-Champaign

Part-Level 3D Generation 3D Shape Generation Text-to-3D Generation Diffusion Models

Open paper AskPDF arXiv

Understanding and generating 3D objects as compositions of meaningful parts is fundamental to human perception and reasoning. However, most text-to-3D methods overlook the semantic and functional structure of parts. While recent part-aware approaches introduce decomposition, they remain largely geometry-focused, lacking semantic grounding and failing to model how parts align with textual descriptions or their inter-part relations. We propose DreamPartGen, a framework for semantically grounded, part-aware text-to-3D generation. DreamPartGen introduces Duplex Part Latents (DPLs) that jointly model each part's geometry and appearance, and Relational Semantic Latents (RSLs) that capture inter-part dependencies derived from language. A synchronized co-denoising process enforces mutual geometric and semantic consistency, enabling coherent, interpretable, and text-aligned 3D synthesis. Across multiple benchmarks, DreamPartGen delivers state-of-the-art performance in geometric fidelity and text-shape alignment.

Abstract

Text-to-3D Generation

Explore similar work

DreamPartGen: Semantically Grounded Part-Level 3D Generation via Collaborative Latent Denoising

ReConText3D: Replay-based Continual Text-to-3D Generation

Arko-T: A Foundation Model for Text-to-Structured 3D Generation