- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes LiTE (LiDAR Trajectory Evolution), a distillation framework for accelerating diffusion-based LiDAR scene completion. LiTE inserts Visibility-Conditioned Latent Transport (VCLT) at a deep sparse U-Net interface and executes several teacher-aligned latent micro-steps within one encoder-decoder pass. VCLT repeatedly updates scan-to-query correspondences, combines a global motion field with local scan guidance, and gates the local contribution using sensor-ray visibility. Training combines EMA endpoint consistency, direct supervision from LiDiff trajectories, intermediate latent-trajectory losses, and terminal structural losses. On various datasets, the four-transition model improves completion metrics over baselines while reporting substantially lower latency.
Critical strengths
-
The method addresses an important computational bottleneck with a coherent LiDAR-specific design. Decoupling expensive sparse U-Net evaluations from lightweight internal trajectory updates is well motivated.
-
The paper reports useful technical details and diagnostics. The appendix specifies the endpoint, intermediate trajectory, free-space, occupancy/density, and local-geometry objectives. Raw and refined outputs are separated, several inference depths are evaluated with one checkpoint, and the paper reports local ray-class diagnostics and the isolated cost of the conditioning component rather than relying only on a single headline result.
Critical weaknesses
- Lack of baseline comparisons and discussions. LiNeXt (2511.10209), LiFlow (2602.02232, RapidLiDAR (2608.16490), etc., are competitive baselines. The relevant methods should be discussed and, ideally, implement under the paper's common protocol for fair comparison.
- The ablations do not isolate the mechanisms used to justify the central novelty. E.g., replacing VCLT with radius attention or closest-point matching simultaneously removes dynamic coordinate transport and visibility routing while changing module placement and aggregation; as acknowledged in the appendix, this comparison does not separately establish the benefit of visibility.
- Several concerns about the efficiency claim. Considering deployment, the current statement of inference efficiency seems insufficient. Besides, the statement that LiTE is trained on only 5.2% of LiDiff's training set should not be interpreted as end-to-end data efficiency because the student relies on a LiDiff teacher that was itself trained on the full dataset.
Questions for authors
- Within VCLT, what are the independent contributions of dynamic candidate refresh, ray-visibility routing, the global motion field, GRU history, and free-space supervision?
- How sensitive is the method to the number and spatial resolution of latent queries? What happens when initially assigned queries cross or move far from their original lattice cells? Was dynamic particle-to-query reassignment tested?
- I would like to see some more comparisons to recent related works such as LiFlow or RapidLiDAR as mentioned above.
- Can the authors provide a parameter- and compute-matched control to show that the gain comes from teacher-aligned latent trajectory supervision rather than additional recurrent depth or model capacity?
- I'm also quite concern the efficency regarding deployment online. Could you clarify the latency protocol, including whether one sample uses one or four GPUs, batch size, warm-up, synchronization, preprocessing, data transfer, and refinement? Any detailed inference protocal would be helpful.
5 comments
Follow-up to Reviewer KDhK
Dear Reviewer KDhK,
Thank you once more for your thoughtful comments, especially on deployment efficiency and the behavior of the latent queries. Since the discussion period will close soon, we would like to gently follow up on our response.
Following your suggestions, we evaluated LiNeXt, RapidLiDAR, and LiFlow under a common protocol and showed that LiTE can distill LiFlow into a stronger four-step student. We also isolated each VCLT component, provided a parameter- and compute-matched control, detailed the full latency protocol, clarified the scope of the 5.2% claim, and explained why moving queries do not cause a mismatch at inference. We believe these cover all the points in your review.
We would greatly appreciate your thoughts on whether these responses address your concerns. If there are further questions or experiments you would find helpful, we would be glad to discuss them before the deadline.
Best regards, The Authors
Opioions on the first round of replies
Thank you for the additional experiments and clarifications. The new ablations address much of my concern about mechanism attribution, and I appreciate the corrections to the efficiency and performance claims.
However, these results also suggest a narrower contribution than originally presented. The end-to-end speedup is modest, and the accuracy advantage is limited to the compared distillation methods. For now I decide to keep my initial rating.
Response to Reviewer KDhK
We thank the reviewer for the detailed questions on deployment and on the latent queries. Note: this reply refers to Table R1 (General Comment, Part 1), Tables R2–R3 (General Comment, Part 2), and Table R4 (General Comment, Part 3); please see these comments for the full tables.
W1 / Q3: LiNeXt, LiFlow, and RapidLiDAR. Following the reviewer's suggestion, we evaluated all three methods from their public weights under our common protocol with the same refiner, which confirms the comparison in Table R2 and also provides IoU for RapidLiDAR. LiNeXt and RapidLiDAR outperform LiTE (CD† 0.149 and 0.138 vs. 0.283). However, they address a different problem: they are regression networks that predict the completed scene in a single forward pass, with no diffusion process involved. LiTE is a distillation method: given a pretrained generative teacher, it trains a few-step student from only 1,000 scans without retraining the generative model, and among distillation methods it achieves the best CD and JSD with a quarter of the backbone evaluations (Table R2). As shown in Table R2, LiFlow has lower CD than LiTE (CD† 0.228 vs. 0.283), while LiTE is better on JSD (JSD† 0.361 vs. 0.367) and fine-scale IoU (IoU† at 0.2 / 0.1 m: 25.81 / 13.67 vs. 25.1 / 12.9). Being a generative model, LiFlow can also serve as a stronger teacher for LiTE, so we distilled it: the four-step student surpasses LiFlow itself on every refined metric while denoising 2.9× faster (Table R3).
W2 / Q1: Independent contributions inside VCLT. We agree that replacing VCLT with Radius Attention or Closest Point removes several components at once. The five components map to single-factor controls in Table R4 (full model: CD† 0.283, JSD† 0.361, IoU† 0.1 m 13.67):
Each component contributes beyond seed variance, with GRU history and free-space supervision mattering most.
W3: Efficiency and the 5.2% claim. As corrected in Table R1, LiTE is 2.8× faster than ScoreLiDAR in denoising (2.3× when averaged over frames) and 1.4× faster end-to-end. On data, we agree that 5.2% should not be read as end-to-end data efficiency, since the teacher was trained on the full dataset. What it does reflect is how lightly LiTE uses the teacher: unlike ScoreLiDAR, whose student is initialized from LiDiff weights and queries the teacher throughout training, LiTE is trained from random initialization and uses the teacher only offline, to generate trajectories for 1,000 scans. The revised paper describes 5.2% as the data budget of the distillation stage.
Q2: Latent queries. The queries are the occupied sites at the deep interface of the U-Net, so their number and resolution follow from the backbone stage rather than a free hyperparameter. At inference there is no particle-to-query correspondence. A query's position is only a lookup point: at every micro-step, scan candidates and visibility are retrieved again from its current position, while its hidden state stays on its own lattice site and is decoded there. A query that moves far from its cell therefore simply reads evidence from its new location without causing any mismatch. The assignment π(n) is used only during training to build supervision targets; dynamic reassignment would change only these targets.
Q4: Parameter- and compute-matched control. Row C1 in Table R4 is this control: the same architecture, micro-steps, parameters (33.70 M), and inference time as the full model, with only the teacher-aligned supervision removed. CD† degrades from 0.283 to 0.473, so the gain does not come from recurrent depth or capacity. VCLT itself adds only 0.30 M parameters.
Q5: Latency protocol for online deployment. Each frame uses one GPU (not four) with batch size 1, and
torch.cuda.synchronize()is called around every timed segment; baseline times are single runs without a separate warm-up, and LiTE's denoising time is the per-frame mean over sequence 08. End-to-end time includes filtering, CPU FPS, host-to-GPU transfer, and post-processing; disk reading is excluded, and the shared refiner adds 0.09 s for every method. For online deployment, the CPU FPS inherited from the LiDiff protocol now dominates (2.97 s of 3.93 s), so the system is not yet real-time; we have added this as a limitation, and removing it requires training with a cheaper sampler.