A spinning LiDAR sensor measures the world along a fixed fan of rays. Each ray has a direction set by the hardware and returns one distance, or nothing. Generative models of LiDAR scenes either ignore this structure, moving free 3D points around in $(x, y, z)$, or they respect it by working on a "range image": a 2D grid indexed by the angles of each beam, where each pixel stores a distance. The range-image models are the most realistic available, but they are slow. They are diffusion or flow models, meaning generators that turn random noise into a sample through dozens of small network evaluations applied one after another. Spiral needs 5.7 seconds per annotated scene on an A6000, and its authors call this too slow for on-the-fly driving simulation. Separately, one-step generators based on "average velocity" now produce good images in a single network call. Mean Flows reports ImageNet 256 FID 3.43 at one evaluation. 3D MeanFlow brings the same idea to point clouds, but only in Cartesian coordinates and only at the scale of single objects.
This note proposes one-step generation of full LiDAR scenes in which ray directions stay fixed, as the sensor fixes them, and the generator moves only a per-ray distance, together with reflectance and a "did this ray return" mask. The point is not merely to run Mean Flows on range images, which is the obvious and nearly trivial combination. The aim is to identify and fix the specific way one-step generation fails in this geometry. A single averaged jump cannot choose between two surfaces that the same ray might hit, for example a near car versus a far wall. The result is points floating in empty space between them, which is exactly the boundary blur that LiDM and R3DPA report for range-image models. We argue that three choices determine how much of this blur appears: how distance is warped before transport, how ray directions are encoded so attention can reason along lines of sight, and whether each ray is allowed to commit to one of several candidate surfaces.
Core Claim
Once the sensor's rays are held fixed, LiDAR scene generation reduces to transporting a scalar distance along each ray, conditioned on neighboring rays. Along each ray, the distribution of possible distances is sharply multimodal exactly at occlusion boundaries. An average-velocity generator that takes one step regresses toward the mean of those modes and places points in empty space.
We claim that a one-step generator with three ingredients can approach multi-step range-view quality on SemanticKITTI and KITTI-360 at under 100 ms per scene, with markedly fewer floating points than a plain Mean Flows range-image model given the same backbone. The first ingredient is a warped radial coordinate, comparing raw range, log range and inverse range. The second is a direction encoding that keeps ray angle separate from content. The third is a lightweight two-hypothesis output per ray.
The new contributions are the diagnosis (that one-step error concentrates in occlusion-induced along-ray bimodality) and a measured answer to which polar parameterization minimizes it. No paper in the file applies average-velocity transport to range-view or polar LiDAR data. The closest work, 3D MeanFlow, is Cartesian and object-level.
Background and How This Differs
Three range-view generators define current quality and also the latency problem. LiDM compresses range images with beam-shaped 1×4 kernels and runs latent diffusion, reaching 1.739 samples per second. R3DPA adapts an ImageNet-pretrained flow-matching transformer (SiT-B/2) to range latents and reports a 17% improvement in its point-cloud distribution distances over R2Flow and LiDM on KITTI-360. Spiral jointly generates depth, reflectance and semantics, and feeds confident labels back into geometry across iterations. All three motivate this proposal: they show that the sensor-aligned representation works, yet each needs many sequential steps. LiDM and R3DPA also admit boundary blur on back-projection. We keep their representation and change the sampler and the output parameterization.
Simultaneous Diffusion Sampling runs several diffusion streams across synthetic viewpoints to remove floating points. It reaches the same symptom we target, but by multiplying latency rather than cutting it. Its novel-view depth error of 2.33 versus 2.36 for single-view sampling on KITTI-360 Drive 0 is a useful reference for how much multi-view consistency buys.
The one-step machinery comes from Mean Flows, which trains a network $u(x_t, r, t)$ to predict the average velocity over an interval $[r, t]$. It uses the identity
$$u(x_t, r, t) = v(x_t, t) - (t - r)\,\frac{d}{dt}u(x_t, r, t),$$
where $v$ is the instantaneous velocity of the noise-to-data path, $r$ and $t$ are the interval endpoints in generation time, and the time derivative is computed with a Jacobian-vector product. 3D MeanFlow supplies two pieces we reuse: a Chamfer shape-consistency term against single-step drift, and the observation that one-step sampling loses diversity. Its Cartesian, permutation-invariant set formulation is exactly what we drop. With rays fixed, there is no permutation problem: point $i$ is always ray $i$.
The polar-encoding literature supplies the inductive bias and the evidence that it matters. PTNet puts $(\rho, \phi)$ into transformer positional encodings so a detector can separate foreground from background along lines of sight. Ego3RT uses an ego-centered polar grid and improves drivable-area IoU from 72.9 to 79.6 over Lift-Splat-Shoot. PoPE shows that rotary embeddings entangle content with position, and that a polar decoupling solves an indirect-indexing task (94.82% versus 11.16%). Both PTNet and Ego3RT are discriminative, so they motivate rather than supply our method. PoPE supplies one candidate direction encoding, applied here to 2D beam angles rather than 1D sequence positions.
One constraint matters. The tangent-normal analysis in Learning Manifold Data with Flow Matching assumes flat data subspaces. Our per-ray warped distance is a flat 1D coordinate, so the analysis applies more naturally here than to curved Cartesian scene manifolds. We treat that only as motivation and not as a guarantee.
The well-known polar singularity at the origin does not arise in our setup: rays start at the sensor and never return range zero. Rays with no return are handled by an explicit mask channel rather than a sentinel distance.
Research Program
Base model. We build a range-view transformer at the scale of Spiral's 61M parameters, operating on 64×1024 (SemanticKITTI) or KITTI-360 range images. The channels are warped range $g(r)$, reflectance, and a return-mask logit. Semantics are predicted by a head on the final output, because Spiral's iterative label feedback has no analogue in a single step. Training follows Mean Flows from scratch, with interval sampling and the identity above, plus a Chamfer term on back-projected points as in 3D MeanFlow.
Radial warp. We compare three choices: $g(r) = r$, $g(r) = \log r$ (log-polar, echoing the log-polar interaction grids in Tracking People and Their Objects), and $g(r) = 1/r$. The warp controls how far apart near and far surface modes sit in the transported coordinate, and so how badly a single averaged step blends them. We do not presuppose a winner. Inverse range compresses distant background, where LiDAR returns are sparse and noisy, but it exaggerates near-field variation. The experiments decide.
Direction encoding. We compare learned absolute embeddings, 2D rotary embeddings on (azimuth, elevation), and a PoPE-style polar embedding in which content magnitudes and angle-dependent phases are kept apart. The hypothesis is that decoupled direction encoding lets attention find "the same surface on the neighboring ray" more reliably, which is what decides which mode a ray should commit to.
Two-hypothesis head. Each ray outputs two candidate warped ranges and a selection logit. Training uses a winner-takes-all loss on the one-step endpoint, with the Mean Flows objective applied to the selected branch. This gives the single step a way to choose a surface rather than average two. It is the mechanism we expect to matter most. It is also the ingredient a skeptic will most want isolated, so every experiment includes it on and off.
Diagnosis tool. For conditional settings with ground truth, we define a "floating-point rate": the fraction of generated returns whose distance falls in the empty gap between two ground-truth surfaces on that ray or its 3×3 neighbors. The conditional settings are beam upsampling from 16 to 64 beams and inpainting of masked azimuth sectors on KITTI-360. We report the floating-point rate stratified by the size of the depth discontinuity and by range bin. If the core claim is right, one-step error, but not multi-step error, should concentrate at large discontinuities.
Baselines
Naive reference: Spiral with fewer steps. The do-nothing reference is the strongest available multi-step range-view model run at its native schedule, then simply truncated. We retrain Spiral on SemanticKITTI from its released recipe at matched parameter count and measure S-FRD, S-FPD, and wall-clock latency on one A6000. We then rerun it with a DPM-Solver-style fast ODE sampler at 4, 2 and 1 steps.
Spiral reports a 31% S-FRD and 56% S-FPD improvement over R2DM plus RangeNet++ and 5.7 seconds per scene. Absolute S-FRD and S-FPD values: no reported result found on Lacuna, so we will reproduce them. Beating this reference means matching full-schedule Spiral within a pre-registered margin (10% relative on S-FRD and S-FPD) at under 100 ms. It also means clearly outperforming truncated Spiral at the same number of network calls.
Strongest existing method: distilled range-view flow. The strongest competitor for "fast range-view generation" is not in the file as a finished system. It is R3DPA's flow-matching range model compressed by progressive distillation to 2 and 4 steps. That is the route a practitioner would actually take. For R3DPA the file gives only a relative figure, the 17% distribution-distance improvement over R2Flow and LiDM on KITTI-360. No absolute FSVD/FPVD numbers were found on Lacuna. For LiDM the file gives a throughput of 1.739 samples per second.
We run the distillation with the same data and a compute budget equal to our full training, and tune the step schedule on the same validation split. Beating it means equal or better FPVD/FSVD at 1–2 network calls, plus a lower floating-point rate.
Coordinate-choice baseline: Cartesian one-step. To test whether sensor-aligned rays matter at all, we train a scene-level Cartesian 3D MeanFlow model: the set-based identity plus Chamfer loss, with points sampled to the same count. Its object-level reported numbers are about 5 ms per shape and over 8× faster than PSF. Its Chamfer values on ShapeNet are described only qualitatively, so no reported number was found on Lacuna. For scene completion, the multi-step reference is LiDPM, at 36.6 voxel IoU at 0.5 m on SemanticKITTI with about 20 DPM-Solver steps. Our one-step model, conditioned on a single scan, should approach that figure in one call.
Ablations. Two ablations isolate the key ingredients. The first is plain Mean Flows on raw range with learned positional embeddings and no two-hypothesis head; the comparison with it is the proposal's core test. The second varies the warp, the encoding, and the head one at a time, all at 1 and 2 network calls.
Why not just…?
Why not just run plain Mean Flows on an existing range-image backbone? This is the real threat, and it is built into the experiments as the main ablation rather than argued away. If plain Mean Flows on raw range matches our full model in FPVD and in floating-point rate at occlusion boundaries, then the polar-specific ingredients add nothing and the contribution reduces to "Mean Flows works on LiDAR." That would still be a useful finding, but it is a smaller one, and we will report it as such. Our reason to expect a gap is the boundary blur already documented in LiDM and R3DPA for multi-step models. One averaged step should make it worse, and the stratified floating-point rate will show whether it does.
Why not just distill Spiral or use a fast ODE solver? LiDPM still needs about 20 solver steps, and Spiral at its schedule takes seconds. Mean Flows reports one-step FID 3.43 against 10.60 for Shortcut models and 34.24 for consistency training on ImageNet, which suggests that training for average velocity from scratch beats squeezing a multi-step model. That evidence comes from images, though, not LiDAR. The distilled-R3DPA and truncated-Spiral baselines settle the question directly at matched compute. If distillation to 2 steps matches us, our case narrows to the 1-step regime.
Why not just generate Cartesian points with 3D MeanFlow? Fixing ray directions removes the permutation problem and the need for set-level matching. It also guarantees the sensor's beam pattern by construction, which LiDM had to engineer into its compressor. Against that, the Cartesian model can place points off-ray, which may help when the target sensor differs from the training sensor. The Cartesian baseline above measures whether any of this shows up in FPVD, in completion IoU, or in the far-range bins where PTNet found Cartesian processing loses sparse geometry.
Why not just use rotary embeddings on the beam angles? PoPE's indirect-indexing result shows that rotary embeddings can fail when content and position must be matched separately. Whether that failure occurs along LiDAR rays is unknown. The encoding ablation answers it. If rotary embeddings perform equally well, we keep them for their speed, since PoPE adds roughly $2L^2d$ FLOPs per layer.
Proposed Experiments
Unconditional and semantic generation run on SemanticKITTI, reporting S-FRD and S-FPD against Spiral. Unconditional generation on KITTI-360 reports FRD, FPVD and FSVD against LiDM, R3DPA and distilled R3DPA. Conditional tasks on KITTI-360, namely beam upsampling, sector inpainting, and novel-view depth following Simultaneous Diffusion Sampling's protocol, report depth error and floating-point rate. Scene completion on SemanticKITTI reports voxel IoU at 0.5 m against LiDPM. Latency is wall-clock on one A6000 at batch size 1, including back-projection. Diversity is measured as coverage and pairwise sample distance, because both 3D MeanFlow and Mean Flows warn that one-step sampling loses it. We report 1, 2 and 4 calls to show how much a second step recovers.
Three results decide the question. First, the full model must beat plain range-view Mean Flows on floating-point rate at large depth discontinuities by a clear margin, while matching it elsewhere. Without that, the diagnosis is wrong. Second, it must match full-schedule Spiral within 10% on S-FRD and S-FPD at under 100 ms. Without that, the latency claim fails. Third, the best radial warp must hold across both datasets. If the best warp flips between datasets, the "which polar coordinate" finding is a tuning artifact rather than a property of LiDAR geometry.
References
- Spiral: Semantic-Aware Progressive LiDAR Scene Generation and Understanding
- Mean Flows for One-step Generative Modeling
- 3D MeanFlow: One-Step Point Cloud Completion and Generation via Average-Velocity Transport
- Towards Realistic Scene Generation with LiDAR Diffusion Models
- R3DPA: Leveraging 3D Representation Alignment and RGB Pretrained Priors for LiDAR Scene Generation
- Simultaneous Diffusion Sampling for Conditional LiDAR Generation
- PTNET: A PROPOSAL-CENTRIC TRANSFORMER NET- WORK FOR 3D OBJECT DETECTION
- Learning Ego 3D Representation as Ray Tracing
- Decoupling The "What" and "Where" With Polar Coordinate Positional Embedding
- Learning Manifold Data with Flow Matching
- Tracking People and Their Objects