OT-MeanFlow3D: Bridging Optimal Transport and Meanflow for Efficient 3D Point Cloud Generation

arXiv:2509.22592 · cs.LG · Submitted 2025-09-26 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "OT-MeanFlow3D: Bridging Optimal Transport and Meanflow for Efficient 3D Point Cloud Generation".

Jane: Here's the gist of the paper. The gist The proposed Optimal Transport–enhanced MeanFlow framework (OT-MF3D) bridges the gap between efficiency and fidelity in 3D point cloud generation by integrating optimal transport-based sampling into a single-step MeanFlow paradigm to better preserve geometric and distributional structure while maintaining constant-time sampling How it works The core idea of OT-MF3D is to couple MeanFlow’s average-velocity objective with mini-batch optimal transport matchings during training.

Tom: This coupling aligns source and target point-cloud distributions, which is crucial in the extreme low-step regime where iterative correction is limited, by supplying "geometry-consistent correspondences that straighten the induced transport and reduce multi-modal ambiguity".

Jane: The method operates using a unified formulation defined as LOTMF(uθ) which combines mini-batch OT flow and the mean flow method.

Tom: This is formalized by the loss function LOTMF(uθ):= EXB0 ∼p,XB1 ∼qE(X0,X1)∼π0,1,th∥uθt(t, r, Xt) − utgt(vt, t, r)2i, where π0,1 is an optimal plan for OT(pB, qB), pB = Law(XB0), qB = Law(XB1).

Paper summary: Tom: So we’re wrapping up this look at OT-MeanFlowthree dee, which is about using optimal transport to make those flow matching models for three dee point clouds faster and better.

Jane: It’s really about taking a technique that needs many steps during generation and forcing it into a single step by using geometry-consistent correspondences from optimal transport.

Lu: The authors are proposing this Optimal Transport enhanced MeanFlow framework, which couples the average velocity objective with mini-batch OT matchings to align source and target point cloud distributions.

Meng: So they're trying to get better shape structure in one go without the iterative correction that usually makes these models so slow at inference.

Lalam: It basically says you can get high-quality point cloud generation and completion with reduced inference latency because you’re using this optimized flow approach.

Tom: The title itself, "Bridging Optimal Transport and MeanFlow," tells you exactly what they did: they bridged the gap between a standard flow model and an optimal transport sampling method.

Jane: That bridge is key because standard flow models often struggle with point-level matching when you’re only allowed one step to generate something new.

Lu: What’s interesting is their training method too, which uses a latent-conditioned training algorithm that separates the shape sampling from the point matching process.

Meng: If they can decouple those two parts, it means the model isn't wasting compute trying to solve point-level transport within every single shape generation step.

Lalam: That decoupling is what lets them achieve this strong speed-quality tradeoff, which is what really matters for deployment.

Jane: So for someone just listening, this paper means we can expect three dee generative models to be much more practical and usable in real-time applications because they cut down on the slow sampling time.

Tom: And the authors show that it improves over existing one-step baselines like PSF, which is a good sign that this isn't just theoretical stuff.

Jane: It moves us closer to having three dee reconstruction tools that are genuinely fast enough for things like robotics or augmented reality applications.

Lu: But we still have to keep an eye on those training costs; the paper claims faster training, but we need to see how scalable that actually is when we push the dataset size way up.

Meng: That’s exactly what I want to know—if this works well on ShapeNet, does it hold up when I feed it a massive industrial scan?

Jane: We'll have to look closely at those scaling questions next because efficiency in training is just as important as speed at inference when we think about putting this into the real world.

Conclusion: Tom: So we’re wrapping up this look at OT-MeanFlowthree dee, which is about using optimal transport to make those flow matching models for three dee point clouds faster and better.

Jane: It’s really about taking a technique that needs many steps during generation and forcing it into a single step by using geometry-consistent correspondences from optimal transport.

Lu: The authors are proposing this Optimal Transport enhanced MeanFlow framework, which couples the average velocity objective with mini-batch optimal transport matchings to align source and target point cloud distributions.

Meng: So they're trying to get better shape structure in one go without the iterative correction that usually makes these models so slow at inference.

Lalam: It basically says you can get high-quality point cloud generation and completion with reduced inference latency because you’re using this optimized flow approach.

Tom: The title itself, "OT-MeanFlowthree dee: Bridging Optimal Transport and MeanFlow for Efficient three dee Point Cloud Generation," tells you exactly what they did: they bridged the gap between a standard flow model and an optimal transport sampling method.

Jane: That bridge is key because standard flow models often struggle with point-level matching when you’re only allowed one step to generate something new.

Lu: What’s interesting is their training method too, which uses a latent-conditioned training algorithm that separates the shape sampling from the point matching process.

Meng: If they can decouple those two parts, it means the model isn't wasting compute trying to solve point-level transport within every single shape generation step.

Lalam: That decoupling is what lets them achieve this strong speed-quality tradeoff, which is what really matters for deployment.

Tom: This framework moves us closer to having three dee generative models that are genuinely fast enough for things like robotics or augmented reality applications because they cut down on the slow sampling time.

Jane: It’s a practical solution for real-time and resource constrained three dee point cloud modeling because it maintains efficiency gains alongside improved accuracy.

Lu: We still have to keep an eye on those training costs; the paper claims faster training, but we need to see how scalable that actually is when we push the dataset size way up.

Elaheh Akbari, Shansita Sharma, Ping He, Ahmadreza Moradipari, Kyungtae Han, Hamed Pirsiavash, Yikun Bai, *Soheil Kolouri

Department of Computer Science, Vanderbilt University · Toyota InfoTech Labs · Department of Computer Science, University of California, Davis · Department of Mathematics, Purdue University

cs.LG

Submitted: 2025-09-26

Updated: 2026-10-17

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 83/100

The gist: The gist The proposed Optimal Transport–enhanced MeanFlow framework (OT-MF3D) bridges the gap between efficiency and fidelity in 3D point cloud generation by integrating optimal transport-based

Key concepts

MeanFlow: MeanFlow is a generative modeling framework that aims for single-step generation by learning an average velocity field. It's used for 3D point cloud synthesis, but its quality is sensitive to the training data's geometry.
Optimal Transport (OT): OT is a mathematical tool used during training to create 'geometry-consistent correspondences.' It aligns the source and target point-cloud distributions by finding an optimal plan, which helps reduce ambiguity in low-step sampling regimes.
LOTMF(uθ): This is the unified formulation combining mini-batch OT flow and the mean flow method. The loss function minimizes a combination of these objectives, ensuring that the learned velocity field respects both distribution alignment (via OT) and average motion (via MeanFlow).
Latent-Conditioned Training: This novel training algorithm decouples shape-level sampling from point-level matching. It allows exact point-level OT within each conditioned shape while performing shape transport in a compact latent space, avoiding expensive nested OT computations.

Terminology

Optimal Transport — 3× in this paper · explained in 5 episodes
Generative Modeling — 1× in this paper · explained in 3 episodes

Summary

The gist The proposed Optimal Transport–enhanced MeanFlow framework (OT-MF3D) bridges the gap between efficiency and fidelity in 3D point cloud generation by integrating optimal transport-based sampling into a single-step MeanFlow paradigm to better preserve geometric and distributional structure while maintaining constant-time sampling

How it works

The core idea of OT-MF3D is to couple MeanFlow’s average-velocity objective with mini-batch optimal transport matchings during training. This coupling aligns source and target point-cloud distributions, which is crucial in the extreme low-step regime where iterative correction is limited, by supplying geometry-consistent correspondences that straighten the induced transport and reduce multi-modal ambiguity.

The method operates using a unified formulation defined as LOTMF(uθ) which combines mini-batch OT flow and the mean flow method. This is formalized by the loss function LOTMF(uθ):= EXB0 ∼p,XB1 ∼qE(X0,X1)∼π0,1,th∥uθt(t, r, Xt) − utgt(vt, t, r)2i, where π0,1 is an optimal plan for OT(pB, qB), pB = Law(XB0), qB = Law(XB1).

Key Components and Contributions

The specific contributions of the work are detailed in several points:

We propose a generative framework for point clouds that combines MeanFlow’s average-velocity objective with mini-batch optimal transport (OT) couplings, enabling high-quality one–step generation.

We leverage OT-based matchings to construct geometry-consistent training pairs, which reduce multimodal ambiguity and improve transport alignment in the extreme low-step regime where iterative correction is limited.

We provide a novel training framework for shape generation and completion, enabling efficient training time without distillation or iterative solvers.

The approach achieves a strong speed–quality tradeoff, delivering competitive or state-of-the-art results in point cloud generation and completion while maintaining constant-time sampling.

Methodology Details

The training process involves several stages that are optimized for efficiency:

Pre-training a lightweight PointNet-based autoencoder on each dataset to reconstruct full point clouds using Chamfer Distance as the loss.
Training a lightweight auxiliary model, the “Context Generator,” which maps a d-dimensional Gaussian vector to a d-dimensional latent vector. This generator is used to condition the final Shape Generation model during inference.
Training the Shape Generation model, which is a residual MLP conditioned on latent vectors from the Context Generator for both conditional shape completion and unconditional shape generation.

Experimental Results

Experiments on the ShapeNet dataset validate the effectiveness of OT-MF3D:

**Our method achieves best or second-best results among one-step generation baselines with comparable inference time, such as PSF[50] and SetVAE[23]. **

Our method is also more efficient to train.

The results show that OT-MF3D achieves significantly improved performance compared to the baselines, including PSF, which is also a flow-based one-step generation method. Furthermore, the model attains the fastest runtime among the compared approaches for shape completion while maintaining high reconstruction quality.

Shape Completion Performance

For shape completion tasks on the GenRe dataset, OT-MF3D demonstrates superior performance:

Our method achieves significantly improved performance compared to the baselines, including PSF, which is also a flow-based one-step generation method.

The results in Table 2 show that OT-MF3D achieves the best CD and EMD, and fastest inference time compared to PSF with one step, and other multi-step baselines. The model can successfully generate the full shape from a partial view in 0.03 seconds.

Ablation Study

An ablation study on the point cloud generation task using ShapeNetChair dataset confirms the contribution of Optimal Transport:

incorporating OT into the MeanFlow algorithm leads to a substantial improvement in performance.

The proposed approach with one step achieves significantly better 1-NNA scores compared to MeanFlow alone, demonstrating that OT provides a meaningful advantage for point cloud generation.

Conclusion

In summary, the paper introduces an Optimal Transport–enhanced MeanFlow framework that successfully bridges the gap between efficiency and fidelity in 3D point cloud generation by integrating optimal transport-based sampling into the single-step MeanFlow paradigm. This approach achieves high-quality generation and completion with reduced inference latency and lower computational overhead during training compared to conventional diffusion and flow-based methods. Evaluations on the ShapeNet dataset validate that OT-MF3D not only improves over existing single-step approaches but also maintains efficiency gains, making it a practical and scalable solution for real-time and resource-constrained 3D point cloud modeling applications.

Improvements for AI systems

Optimal Transport enhanced MeanFlow can generate 3D point clouds in a single step by combining MeanFlow’s average-velocity objective with mini-batch optimal transport matchings to align source and target distributions during training. This allows for high quality generation while preserving constant time sampling, which is crucial for real-time applications like AR/VR.
Geometry consistent training pairs constructed via Optimal Transport couplings reduce multimodal ambiguity and improve transport alignment in the extreme low step regime where iterative correction is limited, leading to higher fidelity and more diverse point clouds.
The model achieves a strong speed quality tradeoff, delivering competitive or state-of-the-art results in point cloud generation and completion while maintaining constant time sampling, making it practical for resource constrained robotic platforms.
For shape completion tasks on datasets like GenRe, the model can successfully capture the full shape from a partial view in very low latency (e.g., 0.03 seconds), enabling real-time reconstruction of 3D objects from sparse inputs.
The framework allows for efficient training, achieving strong performance in fewer epochs compared to conventional flow-based methods like PSF, demonstrating a substantial reduction in training cost while maintaining competitive inference speed.

Abstract

Flow-matching models have recently emerged as a powerful framework for generative modeling, including 3D point cloud synthesis. However, their deployment is limited by the need for multiple sequential sampling steps at inference time. MeanFlow enables single-step generation by learning an average velocity field, but its quality depends heavily on the geometry of the training pairs. In this work, we propose an Optimal Transport-enhanced MeanFlow framework (OT-MF3D) for efficient and accurate 3D point cloud generation and completion. Beyond incorporating OT into MeanFlow, our method introduces a latent-conditioned training algorithm that decouples shape-level sampling from point-level matching. This enables exact point-level OT within each conditioned shape while performing shape-level transport in a compact latent space, avoiding the expensive nested OT computation in point-cloud space. Experiments on standard benchmarks show improved generation and completion quality over fast baselines, with substantially reduced training and inference costs compared with diffusion and other flow-based models.

Sources

Minibatch Optimal Transport and Perplexity Bound Estimation in Discrete Flow Matching