Flow Matching Reinforcement for 3D Mesh Generation via Dynamic Homing Optimization

Zhen Zhou; Zhiwei Ning; Puhua Jiang; Sheng Zhang; Yifei Tang; Jie Yang; Xintong Han; Wei Liu

cs.CVOct 1, 2026

Flow Matching Reinforcement for 3D Mesh Generation via Dynamic Homing Optimization

Authors: Zhen Zhou, Zhiwei Ning, Puhua Jiang, Sheng Zhang, Yifei Tang, Jie Yang, Xintong Han, Wei Liu, +1 more

Organizations: Tencent Hunyuan · Shanghai Jiao Tong University · Communication University of China

Abstract

Flow matching is central to 3D generation, yet in practice its reinforcement learning (RL) methods are largely adapted from 2D visual generation. Representative DPO-, GRPO-, and NFT-style objectives, when applied to negative trajectories, mainly steer predicted velocities away from the corresponding directions without explicitly specifying a target velocity field toward preferred samples. In 3D generation, constrained by pretrained model capabilities, rollout diversity, and reward-distribution complexity, directly applying these RL methods yields limited gains in geometric quality. We introduce a forward-process RL method \textbf{Dynamic Homing Optimization (DHO)}, which reformulates negative-trajectory optimization as positive-sample attraction-guided dynamic homing. Specifically, Minimum-Cost Attractive Matching (MAM) assigns each negative sample a distinct positive target, and Time-Aware Dynamic Correction (TDC) then redirects its trajectory toward the target using a remaining-time-aware corrective velocity. Building on asynchronous online DHO, we develop \textbf{Flow3D-Pro}, an image-to-3D geometry generation framework. Experiments show that DHO outperforms representative DPO-, GRPO-, and NFT-style objectives in 3D generation, while Flow3D-Pro produces higher-quality 3D geometry than existing mesh generation methods.

Figures & tables

Figure 1: We generate diverse 3D assets with plausible global shapes and detailed geometry.

Figure 2: Negative-sample branch objectives across flow matching RL methods.

Figure 3: Method Overview. After pre-training and supervised fine-tuning (SFT), DHO post-trains the rectified flow model in latent feature space using reward-ranked rollout samples. In DHO, MAM assigns each negative sample an attractive positive-target destination, and TDC then constructs a time-aware homing velocity toward the assigned target. We visualize only the negative-sample branch, as the positive-sample branch uses the same MAM rule and homing-velocity construction.

Figure 4: The generated owl gradually becomes canonically aligned during DHO training.

Method	Stage I			Stage II
Method	ULIP $\uparrow$	Uni3D $\uparrow$	User Study $\uparrow$	ULIP $\uparrow$	Uni3D $\uparrow$	User Study $\uparrow$
Base	0.143	0.347	0.60	0.143	0.350	0.71
SFT	0.144 (+0.70%)	0.351 (+1.15%)	1.93	0.143 (+0.00%)	0.352 (+0.57%)	1.91
GRPO-style	0.147 (+2.80%)	0.357 (+2.88%)	3.02	0.147 (+2.80%)	0.358 (+2.29%)	3.00
DPO-style	0.146 (+2.10%)	0.357 (+2.88%)	2.36	0.146 (+2.10%)	0.358 (+2.29%)	2.32
NFT-style	0.146 (+2.10%)	0.355 (+2.31%)	3.07	0.147 (+2.80%)	0.357 (+2.00%)	2.80

Table 1: Comparison of RL objectives for Stage I, evaluated before and after the Stage II refiner.

Figure 5: Qualitative comparison of RL objectives. To facilitate visualization, all generated meshes are rotated to a common orientation. The same convention is used in subsequent visualizations.

Figure 6: Visual comparison with representative 3D generation methods.

Method	ULIP $\uparrow$	Uni3D $\uparrow$	User Study $\uparrow$
Michelangelo Zhao et al. (2023)	0.135	0.290	0.18
Craftsman 1.5 Li et al. (2025b)	0.146	0.325	1.32
Hi3DGen Ye et al. (2025)	0.144	0.335	3.77
Direct3D-s2 Wu et al. (2025b)	0.141	0.345	3.27
Hunyuan3D 2.1 Yang et al. (2025)	0.143	0.347	4.14
Trellis Xiang et al. (2025)	0.145	0.351	4.27

Table 2: Quantitative evaluation on 3D mesh generation.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Illustration of the VLM-based reward evaluation process. The reward model receives a reference image, six canonical-view renderings of the generated mesh, and an evaluation prompt template, then produces criterion-wise assessments and a final scalar reward.

Figure 8: Effect of the target-assignment strategy in DHO.

Figure 9: Effect of positive-sample selection in MAM.

Time Weighting	ULIP $\uparrow$	Uni3D $\uparrow$	User Study $\uparrow$
$w(t)=1$	0.146	0.354	0.45
$w(t)=(1-t)^{2}$	0.149 (+2.05%)	0.359 (+1.41%)	0.55

Appendix

Table 3: Effect of time-aware weighting in TDC.

Velocity Target	ULIP $\uparrow$	Uni3D $\uparrow$	User Study $\uparrow$
Static	0.145	0.353	0.31
Dynamic	0.149 (+2.76%)	0.359 (+1.70%)	0.69

Appendix

Table 4: Static versus dynamic velocity targets for the negative branch.

Advantage Weighting	ULIP $\uparrow$	Uni3D $\uparrow$	User Study $\uparrow$
Without	0.145	0.354	0.42
With	0.149 (+2.76%)	0.359 (+1.41%)	0.58

Appendix

Table 5: Ablation of positive-sample advantage weighting in DHO.

Figure 10: Ablation study on the CFG scale.

$\eta$				$\beta$			ULIP $\uparrow$	Uni3D $\uparrow$	User Study $\uparrow$
0.0	0.1	0.2	0.5	0.0	1.0	5.0	ULIP $\uparrow$	Uni3D $\uparrow$	User Study $\uparrow$
✓					✓		0.145	0.354	1.13
	✓				✓		0.149 (+2.76%)	0.359 (+1.41%)	1.92
		✓			✓		0.147 (+1.38%)	0.357 (+0.85%)	1.68
			✓		✓		0.145 (+0.00%)	0.355 (+0.28%)	1.27
	✓			✓			0.140	0.342	0.30

Appendix

Table 6: DHO performance with the homing coefficient

\eta

and the regularization coefficient

\beta

. The upper block fixes

\beta

and varies

\eta

(where

\eta=0.0

disables the dynamic homing branch for negative-samples), while the lower block fixes

\eta

and varies

\beta

(where

\beta=0.0

removes the reference-policy regularization). Each block is an independent User Study comparison group.

Figure 11: Effect of the rollout group size $G$ on DHO performance.

Rollout Sampler	ULIP $\uparrow$	Uni3D $\uparrow$	User Study $\uparrow$
ODE	0.145	0.353	0.33
SDE	0.149 (+2.76%)	0.359 (+1.70%)	0.67

Appendix

Table 7: Comparison of ODE and SDE rollout samplers.

Figure 12: DHO training dynamics with ODE and SDE rollouts. From left to right: negative-branch gain, positive-branch gain, and reference-policy deviation.

Reconstruction Fidelity	Geometric Plausibility	Orientation Alignment	User Study $\uparrow$
–	–	–	0.63
✓			1.46
✓	✓		1.85
✓	✓	✓	2.06

Appendix

Table 8: Impact of the reward dimensions used for DHO post-training, where “–” denotes the SFT model.

Method	ULIP $\uparrow$	Uni3D $\uparrow$
Michelangelo Zhao et al. (2023)	0.115	0.261
Craftsman 1.5 Li et al. (2025b)	0.129	0.298
Trellis Xiang et al. (2025)	0.126	0.311
Hunyuan3D 2.0 Zhao et al. (2025)	0.130	0.315
Hi3DGen Ye et al. (2025)	0.112	0.299
Direct3D-s2 Wu et al. (2025b)	0.122	0.314

Appendix

Table 9: Quantitative evaluation on the LATTICE-Bench.

Method	ULIP $\uparrow$	Uni3D $\uparrow$	User Study $\uparrow$
SFT	0.144	0.351	0.97
RFT	0.145 (+0.69%)	0.355 (+1.14%)	1.39
PFM Kim et al. (2024)	0.147 (+2.08%)	0.356 (+1.42%)	1.70
DHO	0.149 (+3.47%)	0.359 (+2.28%)	1.94

Appendix

Table 10: Comparison with additional post-training methods.

Figure 13: Visual comparison against commercial models. To facilitate visualization, all generated meshes are rotated to a common orientation.

Figure 14: More qualitative results of Flow3D-Pro across diverse object categories and geometric structures.

Explore similar work

Flow3D-OPD: Multi-Teacher On-Policy Distillation for 3D Geometry Generation with Flow-Matching Diffusion Transformer
Sep 7, 2026Zhiwei Ning, Zhen Zhou, Puhua Jiang +73D Generative Models Uni-Opd
Hierarchical Flow Matching for 3D Point Cloud Generation
Aug 6, 2026Linhao Wang, Qichang Zhang, Ye Su +1Contrastive Flow Matching Point Clouds
Shandong Normal University · University of Macau
MeshFlow: Mesh Generation with Equivariant Flow Matching
Jun 22, 2026Qi Sun, Kiyohiro Nakayama, Jing Nathan Yan +6Mesh Generation Riemannian Flow Matching
City University of Hong Kong, Hong Kong · Stanford University, USA · Cornell Tech, USA +1

Sep 7, 2026, cs.CVJ/K move · Enter open · S save

Flow3D-OPD: Multi-Teacher On-Policy Distillation for 3D Geometry Generation with Flow-Matching Diffusion Transformer

Zhiwei Ning, Zhen Zhou, Puhua Jiang, Xintong Han+6

3D Generative Models Uni-Opd Transformer Architectures

Open paper AskPDF arXiv

Recent image-to-3D generation models built on flow-matching diffusion Transformers (DiT) can produce high-fidelity meshes, yet their post-training strategy remains largely unexplored. There exist several critical bottlenecks in reinforcement learning: the inherent difficulty of defining comprehensive rewards for 3D geometric quality, and the gradient interference that arises when jointly optimizing heterogeneous objectives. Inspired by the practicability of on-policy distillation (OPD) in large language models and image generation, we propose \textbf{Flow3D-OPD}, a two-stage post-training framework that introduces multi-teacher distillation into 3D geometry generation. In the first stage, we utilize the semi-policy to enhance the foundational capability of the pretrained model and then design an agentic verifier for 3D geometric quality evaluation. Based on the verifier, we could cultivate domain-specialized teacher models via direct preference optimization (DPO). In the second stage, we consolidate heterogeneous expertise into a unified student model through on-policy distillation with hard task-routing sampling and gradient accumulation, which could mitigate the gradient interference in joint optimization. Without relying on elaborate modifications, our straightforward yet effective design achieves consistent improvements across all geometric quality dimensions and surpasses all teacher models in the average metric. Extensive experiments demonstrate that our approach provides an effective paradigm for reinforcement learning in 3D generation.

Abstract

Explore similar work

Flow3D-OPD: Multi-Teacher On-Policy Distillation for 3D Geometry Generation with Flow-Matching Diffusion Transformer

Hierarchical Flow Matching for 3D Point Cloud Generation

MeshFlow: Mesh Generation with Equivariant Flow Matching