acceptodds
Withdrawn from ICLR 2027

Efficient Point Cloud Upsampling via Local Flow Matching

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Generative point cloud upsampling becomes increasingly expensive as the requested output point count grows. We present FLUSSO, a flow-matching framework that separates sparse global geometric reasoning from dense local generation. A sparse encoder extracts global context, while a shared velocity field evolves particles within adaptive anchor-centered supports. To train these local flows, sliced optimal-transport flow matching constructs candidate correspondences through one-dimensional projections, scores their transport costs in three dimensions, and softly weights multiple low-cost coupling losses. Complementary multi-step local endpoint distribution matching supervises the terminal geometry reached by differentiable model rollouts. At inference, local predictions are merged and subsampled when necessary, allowing one model to support different output cardinalities. Experiments on PU1K and PUGAN demonstrate competitive reconstruction quality and favorable latency and memory efficiency under the evaluated protocols. At 16× upsampling on PU1K, FLUSSO reduces squared Hausdorff and point-to-surface errors by 42.0% and 26.0%, respectively, compared with PUFM evaluated using the same metric pipeline. A model trained at 4× also generalizes to the evaluated unseen ratios from 5× to 32× without retraining.

Primary area:
generative models
Keywords:
Point Cloud Upsampling, Flow Matching, 3D Point Clouds, Generative Models, Sliced Optimal Transport, Efficient Deep Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_kSiWOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses the computational cost and correspondence ambiguity in flow-matching-based point cloud upsampling. The authors propose FLUSSO, which separates sparse global feature extraction from dense local generation using anchor-conditioned flows. The framework further introduces sliced optimal-transport flow matching (SOT-FM) and multi-step local endpoint distribution matching (LEDM) to improve transport supervision and reconstruction quality. Experiments on PU1K and PUGAN demonstrate competitive reconstruction accuracy, improved inference efficiency, and generalization to unseen upsampling ratios.

Critical strengths
  1. The paper identifies an important scalability limitation of generative point cloud upsampling and proposes a meaningful separation between sparse global reasoning and dense local generation. This design directly addresses the computational bottleneck and supports variable output cardinalities without retraining.

  2. FLUSSO demonstrates improvements in reconstruction quality, particularly in Hausdorff and point-to-surface distances, under controlled comparisons with PUFM. The experiments further examine inference latency, memory consumption, unseen upsampling ratios, and the contributions of the proposed training objectives, providing useful evidence of its practical effectiveness.

Critical weaknesses
  1. PUFM has already addressed point correspondence ambiguity through transport-based pre-alignment, while PUFM++ further introduces endpoint refinement to improve the terminal distribution. Although FLUSSO develops a different local formulation, the proposed SOT-FM and LEDM largely build upon established transport and distribution-matching techniques. The paper does not sufficiently establish the distinctive methodological advantages of these components, particularly without a direct comparison with PUFM++.

  2. The substantial high-ratio speedups are measured mainly against recursively applied PUFM, making it difficult to separate the benefits of local generation from those of avoiding recursive inference. Moreover, the ablations do not directly measure the claimed trajectory mismatch reduction or isolate the architectural contribution of sparse-global/local-dense decomposition from the training objectives. Consequently, the evidence does not fully establish which proposed mechanisms are responsible for the reported gains.

Questions for authors
  1. Since PUFM++ already incorporates endpoint refinement and enhanced flow-matching optimization, could the authors provide a controlled comparison with PUFM++ under identical evaluation protocols, including reconstruction accuracy, inference latency, and memory consumption? This comparison is important for establishing the novelty and practical significance of FLUSSO.

  2. Could the authors provide controlled ablations that isolate the sparse-global/local-dense architecture from SOT-FM and LEDM, preferably with comparable training computation? In particular, can the claimed reduction in trajectory mismatch be directly quantified rather than inferred solely from improvements in final reconstruction metrics?

  3. To what extent do the reported speedups arise from local flow computation rather than avoiding recursive upsampling and global point selection? A comparison that better separates these factors would help establish the intrinsic computational advantages of the proposed architecture.

Reviewer_DvvgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FLUSSO, a point cloud upsampling framework that combines sparse global feature encoding with shared local flow generation. Training uses multiple sliced transport plans for flow-matching supervision and a differentiable multi-step endpoint distribution loss. A single checkpoint supports different output cardinalities through variable particle counts and optional farthest point sampling. Experiments on PU1K and PUGAN demonstrate competitive reconstruction quality and favorable inference latency and memory usage, with additional evaluations on unseen ratios, noisy inputs, and real scans.

Critical strengths
  1. Separating sparse global encoding from local generation is a practical design that supports flexible output cardinalities. The reported latency and memory measurements demonstrate useful benefits, particularly at high upsampling ratios.

  2. The reconstruction results are promising. Shared-pipeline comparisons show improved HD and P2F over PUFM on both datasets. The supplementary candidate-selection ablations also help clarify the role of FPS.

Critical weaknesses
  1. The methodological advance over PUFM is insufficiently established. PUFM already applies flow matching to upsampling and addresses correspondence ambiguity through EMD pre-alignment. FLUSSO adds a local architecture, multi-plan sliced coupling, and endpoint supervision, but the ablations do not adequately establish their independent contributions. In particular, there is no comparison with EMD/three-dimensional OT within the same local architecture, despite using only 24 particles per support. The top-K comparison also uses more training computation than top-1.

  2. The evidence does not fully establish the significance of the proposed combination. The architecture is not separately ablated, and the endpoint experiments do not isolate the benefit of multi-step rollout over simpler endpoint supervision. The largest speedups depend on recursive PUFM inference; the matched single-stage 4× comparison yields approximately 1.3× speedup. Although these gains are useful and the protocol is disclosed, the missing experimental comparison with the discussed PUFM++ baseline further limits the assessment of the advance.

Questions for authors
  1. How does SOT-FM compare with EMD/three-dimensional OT under the same local architecture, in reconstruction quality and training cost? Does top-K retain its advantage at matched training compute?

  2. Can controlled ablations isolate the benefits of the sparse-global/local-generation decomposition and multi-step endpoint supervision?

  3. Why is PUFM++ absent from the experiments, and can a shared-pipeline comparison be provided?

Reviewer_ka6mOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a flow-matching framework named FLUSSO for efficient point cloud upsampling. The method separates sparse global geometric reasoning from dense local generation. A sparse encoder extracts global context, while a shared velocity field evolves particles within adaptive anchor-centered supports. Specifically, the paper introduces Sliced Optimal-Transport Flow Matching (SOT-FM) to construct source–target correspondences through one-dimensional projections and three-dimensional transport costs. It also proposes Multi-Step Local Endpoint Distribution Matching (LEDM) to match the terminal local distribution reached by model rollouts. Experiments on two datasets PU1K and PUGAN demonstrate good results.

Critical strengths
  1. The proposed framework separates sparse global encoding from dense local generation. It avoids global attention over dense candidates and supports different output cardinalities with shared parameters. This design makes the generation of local flows efficient.

  2. The proposed method demonstrates favorable latency and memory efficiency. At 4× upsampling, FLUSSO achieves a 1.3× speedup and a 34.0% reduction in peak memory compared with PUFM. At higher upsampling ratios, the speedup increases significantly under the evaluated recursive PUFM protocol.

Critical weaknesses
  1. The paper compares the proposed method with PUFM, but does not include quantitative comparisons with PUFM++ (PUFM++ is mentioned in the paper). PUFM++ extends PUFM with two-stage training, adaptive time scheduling, and on-manifold sampling constraints. Since both methods are based on flow matching for point cloud upsampling, a comparison with PUFM++ would help better demonstrate the effectiveness of the proposed method.

  2. The main contribution of the paper is to separate sparse global geometric reasoning from dense local generation. However, the ablation studies mainly focus on SOT-FM, LEDM, and candidate selection. There is no direct comparison between the proposed local flow architecture and a global flow architecture under the same training and inference settings. This makes it difficult to determine how much improvement comes from the proposed local generation design.

Questions for authors
  1. Could the authors give some discussions about why is PUFM++ not included in the quantitative comparisons? Since PUFM++ is closely related to the proposed method, could the authors provide a comparison with PUFM++ to better demonstrate the effectiveness of FLUSSO?

  2. Could the authors provide some experimental results comparing the proposed local flow architecture with a global flow architecture under the same training and inference settings? This would help clarify how much improvement in reconstruction quality and computational efficiency comes from the proposed local generation design.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.