{"id":"d7b841ff-cb0f-416f-a6dd-1d8ad36c7115","arxiv_id":"2412.02241","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"R2Flow applies rectified flows with a domain-adapted hourglass Transformer to generate LiDAR range and reflectance images, achieving strong quality with as few as 1 to 4 function evaluations.","lead":"This paper presents R2Flow, a generative model that creates LiDAR range and reflectance images using rectified flows instead of slow diffusion steps. It reports comparable or better sample quality with far fewer sampling steps, which could make LiDAR simulation and data augmentation more practical for robotics.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central few-step advantage is not fully established because the reflow/distillation trajectories are admitted to be non-straight for raydrop pixels; no raydrop-specific quality or curvature check is reported.","rationale":"The reader's weakest-assumption analysis is on target: the paper's contribution is few-step generation, and the straightness of the learned flow is the property that makes few steps work. The authors themselves flag raydrop pixels as a source of curvature in Sec. V and Fig. 8, which is exactly the kind of explicit limitation that should be weighed. My read does not overturn the conditional acceptance: the quantitative results at few steps are substantially better than the tested baselines, and the code and pretrained weights are promised, which supports reproducibility. However, the absence of raydrop-specific diagnostics leaves a real gap between the headline 'outperforms' claim and the evidence. Secondary issues, such as single-run metrics without confidence intervals and the unbenchmarked RangeLDM baseline, reinforce the need for conditionality but are less central than the raydrop-straightness question. The proposed test would settle whether the few-step advantage holds in the most difficult and safety-relevant part of LiDAR data, namely missing returns.","tokens_in":11306,"tokens_out":9521,"duration_ms":102772,"concrete_test":"Compute the Eq. (6) curvature s(t) separately for raydrop pixels (value -1 at t=1) and ordinary pixels in the 2-RF and 2-RF+1-TD models, and add a raydrop-mask JSD metric to the NFE=1 and NFE=4 rows of Table I by comparing the distribution of raydrop masks in 10k generated samples against the real KITTI-360 masks. If raydrop-pixel curvature at the distilled timesteps remains much higher than non-raydrop curvature, and if raydrop-mask JSD degrades sharply at few steps relative to 256 steps, the paper should explicitly qualify the few-step fidelity claim to non-raydrop structure or adopt a raydrop-aware loss.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing condition is that after reflow and timestep distillation, the learned ODE trajectories are straight enough that one to four Euler steps preserve LiDAR fidelity. The authors' own Sec. V states that 'the raydrop pixels drifting toward a value of -1 may hinder the training of straight flows,' and Fig. 8 shows that the most curved trajectories are associated with raydrop pixels. This matters because the few-step models in Table I are built by distilling 2-RF, and 2-RF at 256 steps is itself worse than 1-RF on FRID, FPD, FPVD, and FSVD (8.64 vs 4.18, 11.06 vs 9.32, 29.08 vs 16.51, 29.20 vs 15.07), indicating that the reflow trade-off sacrifices fidelity for straightness. The paper reports no raydrop-specific evaluation: no per-pixel curvature for raydrop versus non-raydrop pixels in 2-RF or 2-RF+TD, no distributional match of raydrop masks, and no error bars. If raydrop-induced curvature persists at the distilled timesteps, the few-step FRD of 336.62 (1-TD) likely reflects degraded handling of missing-return pixels, and the claim of high-fidelity few-step generation would need qualification. The concern is not that rectified flows are invalid, but that the specific data modality, with a mass of -1 pixels, may break the straightness assumption exactly where the paper's selling point lives.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents R2Flow, a generative model for LiDAR range and reflectance images based on rectified flows. The authors train a 1-rectified flow (1-RF) with conditional flow matching, straighten it via reflow to obtain 2-RF, and then apply timestep distillation to produce few-step models (1-TD, 2-TD, 4-TD). They also introduce a pixel-space hourglass Vision Transformer architecture with circular sliding-window attention and landscape-shaped patches. On unconditional generation with the KITTI-360 dataset, R2Flow shows a better FRD-versus-NFE tradeoff than the tested GAN and diffusion baselines, with notably lower FRD at 1-4 function evaluations, while at 256 NFE it is competitive with but not uniformly better than R2DM on all metrics.","tokens_in":11598,"tokens_out":6007,"duration_ms":57405,"significance":"If the results are reproducible, the paper offers a practical speed-quality improvement for LiDAR generation, a domain where diffusion sampling cost is a practical barrier. The application of rectified flows is well grounded, the code and pretrained weights are provided, and the architecture ablation (Table II) supports the design choices in terms of FRD. The main contributions are the engineering of rectified-flow training for LiDAR data and an efficient Transformer backbone. However, the headline claim of outperforming state of the art for both large and small numbers of steps is stronger than the evidence, and the paper does not resolve the acknowledged difficulty of raydrop pixels in learning straight trajectories, which is exactly where the few-step promise is most at risk.","major_comments":[{"comment":"The introduction states that 'our approach outperforms the state-of-the-art results for both large and small numbers of steps,' but Table I does not support the 'large steps' part. At 256 NFE, R2Flow 1-RF is worse than R2DM on FRID (4.18 vs 3.70), FPD (9.32 vs 3.79), FPVD (16.51 vs 10.90), and FSVD (15.07 vs 9.11), and 2-RF is worse on FRD, FRID, FPD, FPVD, and FSVD. The defensible conclusion is that R2Flow offers a better speed-quality tradeoff, not a categorical improvement at high step counts.","section":"Section I and Table I"},{"comment":"The central claim of the paper is that reflow and timestep distillation make trajectories straight enough for one to four Euler steps to preserve LiDAR fidelity. The authors themselves note in Section V that 'the raydrop pixels drifting toward a value of −1 may hinder the training of straight flows,' and Figure 8(b) associates the most curved trajectories with raydrop pixels. Yet no raydrop-specific evaluation is provided: there is no per-pixel curvature comparison between raydrop and non-raydrop pixels for 2-RF or the distilled models, no distributional comparison of raydrop masks in generated versus real scans, and no error bars or repeated-run statistics for any metric. Moreover, even the best few-step result, 2-RF+4-TD, has FRD 187.10, well above the 256-step 1-RF score of 122.81, so 'high-fidelity' few-step generation needs quantitative qualification. Without these analyses, the few-step advantage—the paper's main selling point—remains insufficiently supported for the raydrop component of LiDAR data.","section":"Section V and Figure 8"}],"minor_comments":[{"comment":"The meaning of the 'Output' column is unclear from the table as presented; please clarify whether it indicates the modalities generated (range only, reflectance only, or both) and how each baseline is configured.","section":"Table I"},{"comment":"Table II reports only FRD for the architecture comparison; reporting a subset of the other metrics from Table I would strengthen the claim that the HDiT-based architecture is better overall, not merely on FRD.","section":"Table II"},{"comment":"The paper does not explain why RangeLDM [8], which is discussed in the related work, is omitted from the experimental comparison; a sentence clarifying availability or scope would be helpful.","section":"Section IV-A"},{"comment":"The caption for Figure 8(b) would benefit from explicitly stating that the pixel value −1 at t = 1 corresponds to raydrop noise, and from quantifying how many of the top-200 curved trajectories are raydrop pixels.","section":"Figure 8"},{"comment":"There are minor typographical issues, such as the capitalization after 'Following prior work' in Section III-B and the phrase 'example trajectory' in Section III-A, which should be 'example trajectories'.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is a revision of earlier work by some of the same authors (R2DM is cited as [4] and is by Nakashima and Kurazume). The comparison with R2DM is therefore a self-comparison; while this is not improper, the editor may want to ensure that the baseline is given a fair configuration. The code release is a positive factor. The main technical concern is the unaddressed raydrop issue in the few-step regime, which should be resolved before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent empirical paper that transfers rectified flows to LiDAR range/reflectance generation, with thoughtful architecture work (circular window attention, landscape patches, beam-angle RoPE, APE) and a real speed-quality tradeoff. The few-step numbers are genuinely better than diffusion baselines: at 4 NFE, R2Flow's FRD is 187 vs ~3000 for R2DM at 1 step, and it roughly matches R2DM at 256 steps. That is the headline result and it holds up in the table.\n\nWhat it does well: the authors adapt HDiT sensibly to the panoramic structure, they measure trajectory curvature directly, and they release code and weights. They also improve LiDM with APE for a fairer baseline, which is good practice. The paper is clearly written and honest about limitations—the Sec. V note that raydrop pixels near -1 may hinder straight-flow training is exactly the right caveat.\n\nSoft spots, in proportion: first, the abstract and intro claim 'outperforms the state-of-the-art for both large and small numbers of steps.' Table I does not support that at large NFE: 1-RF at 256 steps is worse than R2DM on FRID, FPD, FPVD, and FSVD, and only slightly better on FRD. The defensible claim is a better speed-quality tradeoff. Second, no error bars or repeated runs; all metrics are single evaluations with 10,000 samples, so differences like 122.81 vs 124.49 in Table II are within noise. Third, RangeLDM is cited but not benchmarked—unfortunate since it is the closest recent LDM baseline. Fourth, the reflow step 2-RF scores worse than 1-RF at 256 steps on most metrics, which the authors attribute to the parent-model upper bound; that is fine, but it means the straightness gains come at a quality cost that should be stated up front.\n\nThe stress-test concern about raydrop curvature is legitimate: Fig. 8b shows the most curved trajectories cluster near -1, and no raydrop-specific metric is reported. But the empirical few-step numbers are still far better than the baselines, so I don't see this as a load-bearing flaw—more a missing analysis that a good revision should add, such as per-pixel curvature for raydrop vs non-raydrop pixels, or raydrop-mask distribution checks.\n\nWho it is for: researchers working on LiDAR generation for simulation or data augmentation, and anyone applying flow matching to structured, sparse sensor data. It deserves a serious referee and should go to review with requests for repeated runs, a softened claim, and a RangeLDM comparison if feasible. I would cite it as the first rectified-flow LiDAR baseline.","headline":"R2Flow brings rectified flows to LiDAR generation with real few-step speed gains, but the 'outperforms' claim outruns the table and the raydrop-straightness question is honestly flagged but unresolved.","tokens_in":12185,"tokens_out":2286,"would_cite":true,"duration_ms":22641,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that rectified flows with reflow and timestep distillation let LiDAR range-reflectance generation keep fidelity at 1-4 sampling steps, matching or beating diffusion baselines at a fraction of the compute.","keywords":["LiDAR generation","rectified flows","flow matching","diffusion models","range images","reflectance images","hourglass transformer","KITTI-360"],"falsifier":"Measure the FRD gap between 2-RF + 4-TD at 4 steps and at 256 steps separately on range-image regions with and without raydrop pixels (pixels whose training target is $-1$); if the gap is concentrated in raydrop regions, or if training a raydrop-aware reflow loss (masking those pixels) does not reduce few-step FRD, the straightness assumption fails.","tokens_in":11078,"feed_emoji":"📡","tokens_out":6766,"duration_ms":65445,"temperature":0.7,"pith_summary":"This paper presents R2Flow, a generative model for LiDAR range and reflectance images built on rectified flows instead of diffusion. The central claim is that rectified flows learn near-straight trajectories between noise and data, so the trained ODE can be sampled accurately with just one to four neural-network evaluations, whereas diffusion models need hundreds. The paper trains a velocity field with conditional flow matching, applies one reflow pass to straighten trajectories, and optionally distills the model to a fixed small number of timesteps. On the KITTI-360 dataset, it reports that the reflowed model with 4-step timestep distillation reaches a Frechet range distance of 187.10, a much smaller degradation from its 256-step score of 148.09 than any diffusion baseline shows, while keeping competitive quality at high step counts. It also shows that its hourglass-transformer velocity estimator beats CNN backbones in both fidelity and compute for this pixel-space task.","feed_headline":"Straight-flow LiDAR model beats diffusion at 1-4 steps","feed_subtitle":"R2Flow's few-step quality nearly matches its 256-step score, unlike diffusion baselines that collapse.","key_machinery":"Rectified flow with reflow and timestep distillation: the model learns a velocity field $v_\\theta(x_t,t)$ by matching the displacement $x_1-x_0$ along linear interpolation paths, then one reflow pass re-trains it on pairs $(x_1,x_0)$ generated by solving its own ODE, straightening trajectories, and a final timestep distillation trains only at the timesteps used by a fixed $k$-step Euler sampler. The velocity estimator is an hourglass transformer (HDiT-style) with circular sliding-window self-attention in the horizontal direction, landscape $1\\times 4$ patches, beam-angle-conditioned rotary position embeddings, and a learned absolute positional bias; these modifications adapt a pixel-space transformer to panoramic LiDAR range and reflectance images. This machinery carries the paper's speed-quality argument: straight trajectories are what make few-step Euler sampling accurate, and the architecture is what keeps the pixel-space cost low enough to make iterative generation practical.","core_discovery":"The paper's central claim is that LiDAR range-reflectance generation can shed the iterative cost of diffusion without losing fidelity by replacing stochastic curved SDE trajectories with deterministic straight ODE trajectories. Concretely, it shows that a 1-rectified-flow model (1-RF) trained with the conditional flow matching loss, after one reflow pass (2-RF) and timestep distillation (k-TD), produces samples whose quality degrades only mildly as the number of Euler steps drops from 256 to 1. The reported numbers are FRD 148.09 for 2-RF at 256 steps, 212.08 at 2 steps with 2-TD, and 187.10 at 4 steps with 4-TD, compared with R2DM's 2981.89 at 1 step and LiDM's 191.00 at 1 step, showing a flatter speed-quality curve than the diffusion baselines. At the high-step end, 1-RF achieves FRD 122.81 at 256 steps, competitive with R2DM's 154.11, so the rectified-flow formulation does not trade away top-end fidelity. The paper attributes the few-step robustness to trajectory straightening and to the proposed transformer architecture, which uses circular sliding-window attention, landscape patches, beam-angle-conditioned rotary embeddings, and an absolute positional embedding to keep azimuth alignment.","pith_inferences":["A testable extension the authors leave implicit: masking or reweighting the flow-matching loss at raydrop pixels (those flowing toward $-1$) before reflow could straighten the remaining curved trajectories and push the 1-2 step scores even closer to the 256-step scores.","Because the 2-RF quality ceiling is set by the parent 1-RF, any gain in the 1-RF model (better architecture, more data, better raydrop handling) should propagate directly to the few-step distilled models, so architecture work and distillation work compound.","The same straightening recipe should transfer to other range-sensor modalities (radar, sonar, depth cameras) that share the raydrop and missing-measurement phenomenon, since the paper's data representation only assumes a spherical range-reflectance image.","If latency rather than FLOPs is the binding constraint in robotics, the paper's comparison suggests that step count, not parameter count, is the right optimization target for LiDAR generative models."],"forward_implications":["If the central claim holds, LiDAR generative priors can run in a handful of network calls, making them usable in onboard robotics loops for sparse-to-dense completion or anomaly detection rather than offline only.","The flat FRD-versus-NFE curve implies that a single trained model can be traded between quality and latency by choosing the number of steps at inference, without retraining.","The inversion capability, demonstrated by scene interpolation, becomes a practical tool for LiDAR scene manipulation because the ODE can be run backward in few steps.","Diffusion-based LiDAR generators should be revisited: the same evaluation protocol now shows rectified flows matching or beating DDPM and SMLD baselines at both high and low step counts.","The architecture results indicate that pixel-space generation with transformers is competitive with latent-space approaches, so blur introduced by autoencoder decoding can be avoided."],"supporting_citations":[{"why":"Introduces rectified flows with reflow and timestep distillation, the core method for straight trajectories and few-step sampling.","marker":"[12]"},{"why":"Supplies the pseudo-Huber loss and U-shaped timestep distribution used to train the 2-RF model.","marker":"[13]"},{"why":"HDiT hourglass diffusion transformer, the base architecture the paper adapts into its velocity estimator.","marker":"[16]"},{"why":"R2DM, the pixel-space DDPM baseline the paper compares against and whose continuous-time setting it follows for data representation.","marker":"[4]"},{"why":"LiDARGen, the SMLD baseline and source of the evaluation metrics and data split protocol used for KITTI-360.","marker":"[5]"},{"why":"LiDM, the latent diffusion baseline that the paper improves with absolute positional embedding and compares at both high and low NFE.","marker":"[6]"},{"why":"KITTI-360 dataset, the sole training and evaluation data for the unconditional generation experiments.","marker":"[17]"},{"why":"The ODE solver used to generate reflow training pairs and to produce the reported samples.","marker":"[33]"}],"fun_headline_variants":["Rectified flows slash LiDAR generation steps without fidelity loss","R2Flow: few-step LiDAR generation with straight ODE trajectories","Rectified flow model matches diffusion quality using 1-4 sampling steps","Fast LiDAR synthesis via rectified flows and efficient transformer"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The few-step advantage rests on the assumption that after reflow and timestep distillation the learned trajectories are straight enough for Euler sampling to stay accurate, and the paper itself notes in Section V that raydrop pixels drifting toward $-1$ may hinder straight-flow training; if those pixels keep trajectories curved, the 1-4 step gains would shrink.","fun_headline_variants_meta":{"raw":{"variants":["Rectified flows slash LiDAR generation steps without fidelity loss","R2Flow: few-step LiDAR generation with straight ODE trajectories","Rectified flow model matches diffusion quality using 1-4 sampling steps","Fast LiDAR synthesis via rectified flows and efficient transformer"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000261,"raw_usage":{"total_tokens":1609,"prompt_tokens":980,"completion_tokens":629,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":556}},"tokens_in":596,"tokens_out":629,"duration_ms":7723,"temperature":1.0,"reasoning_tokens":556,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:41:22.243220+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the FRD gap between 2-RF + 4-TD at 4 steps and at 256 steps separately on range-image regions with and without raydrop pixels (pixels whose training target is $-1$); if the gap is concentrated in raydrop regions, or if training a raydrop-aware reflow loss (masking those pixels) does not reduce few-step FRD, the straightness assumption fails.","supporting_citations":[{"cited_title":"Flow straight and fast: Learning to generate and transfer data with rectified flow,","cited_arxiv_id":null,"evidence_quote":"Introduces rectified flows with reflow and timestep distillation, the core method for straight trajectories and few-step sampling."},{"cited_title":"Improving the training of rectified flows,","cited_arxiv_id":null,"evidence_quote":"Supplies the pseudo-Huber loss and U-shaped timestep distribution used to train the 2-RF model."},{"cited_title":"Scalable high-resolution pixel-space image synthesis with hourglass diffusion transformers,","cited_arxiv_id":null,"evidence_quote":"HDiT hourglass diffusion transformer, the base architecture the paper adapts into its velocity estimator."},{"cited_title":"LiDAR data synthesis with denois- ing diffusion probabilistic models,","cited_arxiv_id":null,"evidence_quote":"R2DM, the pixel-space DDPM baseline the paper compares against and whose continuous-time setting it follows for data representation."},{"cited_title":"Learning to generate realistic LiDAR point clouds,","cited_arxiv_id":null,"evidence_quote":"LiDARGen, the SMLD baseline and source of the evaluation metrics and data split protocol used for KITTI-360."},{"cited_title":"Towards realistic scene generation with LiDAR diffusion models,","cited_arxiv_id":null,"evidence_quote":"LiDM, the latent diffusion baseline that the paper improves with absolute positional embedding and compares at both high and low NFE."},{"cited_title":"KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2D and 3D,","cited_arxiv_id":null,"evidence_quote":"KITTI-360 dataset, the sole training and evaluation data for the unconditional generation experiments."},{"cited_title":"torchdiffeq,","cited_arxiv_id":null,"evidence_quote":"The ODE solver used to generate reflow training pairs and to produce the reported samples."}],"review_version":1}