What does 'real-time' actually mean for LiDAR scene completion?
For automotive LiDAR, real-time means keeping up with the sensor's acquisition rate—typically 10 Hz, or one full scene every 0.1 seconds. The fastest method in these studies, RapidLiDAR, hits exactly that: it completes a full scene in 0.1 seconds, which is 2.3 times faster than the previous fastest approach [3]. That's the practical speed boundary for a self-driving stack that needs to react to the world as it happens.
But speed alone isn't safety. A 2024 diffusion-based method, LiDiff, produces high-quality completions but takes 30.55 seconds per frame before distillation, and even after optimization it's 5.37 seconds—nowhere near real-time [1]. So the boundary isn't just 'fast enough'; it's 'fast enough while still being accurate enough to trust.' The 0.1-second mark is the target, but only a few methods reach it, and they do so by trading off some completion quality.
Where should you draw the safety boundary? It depends on the type of missing data.
The key insight from a controlled 2026 study is that the safety boundary isn't a single line—it's a set of lines that depend on what kind of gap you're trying to fill. The study compared an iterative refiner (which spends extra compute refining a coarse prediction) against a wider one-shot predictor (which uses more parameters but only predicts once). When the missing data was a coherent gap—like a contiguous angular region removed from the scan—the iterative system improved mean intersection-over-union (mIoU, a measure of how well predicted geometry overlaps ground truth) by 0.911 points over the wide control, with a confidence interval that cleared a predeclared 0.5-point practical margin [2]. That means for large, contiguous holes, spending extra compute on refinement is worth it.
But for independent thinning—where 75% of points are randomly removed—iteration added only 0.300 points, while augmenting the training data with more varied observations added 5.975 points [2]. So if your sensor produces sparse, scattered returns, the safety boundary is on the training side, not the inference side. And for additive clutter—spurious points that shouldn't be there—neither approach helped at all [2]. This tells you that a real-time system needs to know what kind of degradation it's facing; a one-size-fits-all safety boundary will fail in at least one scenario.
What should a system designer actually do?
If you're building a real-time LiDAR completion system, the evidence suggests you should start with a fast, one-shot predictor that runs at 0.1 seconds per frame, like RapidLiDAR [3]. Then, if you know your sensor data has coherent gaps (e.g., occlusions from large objects), add a fixed-depth refinement step—but only if you have the compute budget. The iterative system in [2] cost 10.74 ms and 0.75 GiB per frame, versus 6.25 ms and 0.23 GiB for the wide control—so refinement is roughly 70% slower and uses 3x more memory. That's a real trade-off for an embedded system.
For sparse, scattered data, the better investment is in training coverage—exposing the model to more varied partial scans—rather than in extra inference compute [2]. And for additive clutter, you need a different robustness mechanism entirely, like outlier rejection or sensor fusion, because neither refinement nor wider models fix it [2]. The bottom line: draw your safety boundary based on the dominant failure mode of your sensor and environment, not on a universal speed target. The studies here agree that real-time is achievable, but they also show that 'real-time' and 'safe' are not the same thing—you have to choose where to spend your compute and training resources.
About These Sources
This answer is built on 5 studies (3 peer-reviewed, 2 preprints) — published from 2023 to 2026, 4 from 2024 or later — selected as the most relevant from 7 studies that passed quality screening, drawn from 52 papers retrieved from a database of over 500 million.
Sources used in this answer
Distilling Diffusion Models to Efficient 3D LiDAR Scene Completion
ScoreLiDAR distills a diffusion model for LiDAR scene completion, cutting per-frame time from 30.55 to 5.37 seconds (>5x speedup) on SemanticKITTI while improving quality, but still not real-time.
Iterate or Widen? When Test-Time Refinement Helps LiDAR Scene Completion: A Controlled Study of Evidence Geometry, Training Coverage, and Compute
In a controlled study across 815 SemanticKITTI frames, iterative refinement beat a wider one-shot predictor by 0.911 mIoU points for coherent gaps, but only 0.300 points for 75% thinning, while training augmentation added 5.975 points; neither fixed additive clutter, and refinement cost 10.74 ms and 0.75 GiB per frame.
Towards Real-Time and Adaptable LiDAR Scene Completion
RapidLiDAR completes a full scene in 0.1 seconds (2.3x faster than prior fastest) on SemanticKITTI and KITTI-360, matching 10 Hz LiDAR, with performance on par with state-of-the-art.
Scaling Diffusion Models to Real-World 3D LiDAR Scene Completion
LiDiff applies diffusion directly to point clouds for scene-scale completion, producing more detailed scenes than prior methods but without real-time performance.
SSC-RS: Elevate LiDAR Semantic Scene Completion with Representation Separation and BEV Fusion
SSC-RS uses separate semantic and geometric branches with BEV fusion to achieve state-of-the-art performance on SemanticKITTI while running in real-time.