Abstract
Generative discrete diffusion recasts outdoor LiDAR semantic scene completion via paired sparse-dense synthesis, semantic-guided generation, and structured refinement to achieve state-of-the-art results.
Outdoor LiDAR semantic scene completion (SSC) recovers a dense semantic voxel grid from a scan observing 1% of the target volume, under class imbalance beyond 7,000x. We recast SSC as generative semantic scene completion (GSSC): a single discrete-diffusion formulation in three roles. First, paired sparse-dense scene synthesis (PS^3) generates matched sparse LiDAR observations with their dense semantic completions, addressing the long tail at its source and yielding the PS^3-SemanticKITTI corpus we train on alongside SemanticKITTI. Second, semantic-guided generative scene completion (SGSC) generates the scene from noise with multinomial discrete diffusion, conditioned on the sparse scan through a bird's-eye-view semantic map and a sparse 3D feature stream. Third, the same framework instead refines an existing completion in one flow-matching step: structured source discrete diffusion (S^2D^2). S^2D^2 improves the mIoU of SGSC's own output and every external SSC base tested, without base retraining or test-time adaptation. On the strongest base, one step without test-time augmentation reaches 38.8% mIoU on the SemanticKITTI hidden test. To our knowledge that is the best causal, single-sweep, single-sample result on that leaderboard, +2.1 pp over the previous best published score under the same restriction. Four correction steps with eight-view test-time augmentation reach 39.2%, outside that restriction.
Community
GSSC replaces a frozen scene-completion model's argmax with one cheap correction step. The base network (SCPNet) is left untouched; we run a single discrete-diffusion refinement pass on the probability simplex over its output. No retraining of the base, no distillation, no test-time adaptation.
- 38.8% mIoU on the SemanticKITTI hidden test set, single sweep, N=1, no test-time augmentation. To our knowledge the best causal, single-sweep, single-sample result on the leaderboard, and +2.1 pp over the previous best published score under that restriction (SCPNet, 36.7).
- +2.36 pp on val seq 08: 36.17 -> 38.54 mIoU with one correction step. Three commands reproduce that number in about six minutes on one H100 once the base predictions are local.
- We also report 39.2% mIoU / 59.0 IoU_cmpl with four correction steps and 8-view D4 TTA. Codabench shows each team's best entry, so that is the row visible there, but it sits outside the single-sample setting and the +2.1 pp margin above is not read off it. The per-panel IoU numbers in the figures posted with this submission are from that four-step configuration, not from the N=1 headline.
- The correction pass costs 107.2 ms: 9.33 FPS for the pass, 3.23 FPS end to end.
The paper also covers SGSC, from-noise generative completion at 30.5% val mIoU.
- Code: https://github.com/BillyChern/GSSC-S2D2
- Project page: https://shichen.world/GSSC-project-page/
- Checkpoints: https://huggingface.co/Stone-Chern/GSSC-S2D2-checkpoints
- PS3 synthetic scenes: https://huggingface.co/datasets/Stone-Chern/PS3-SemanticKITTI
- Base predictions: https://huggingface.co/datasets/Stone-Chern/GSSC-S2D2-datasets
- Leaderboard: https://www.codabench.org/competitions/13814/#/results-tab
Author here, happy to answer questions.
Two things the abstract compresses that are worth stating plainly:
The 38.8 headline is one correction step, one sweep, no ensembling. That
restriction is what the +2.1 pp is indexed on. Codabench displays each team's
best entry, so the row visible there is the four-step + 8-view D4 TTA one at
39.2 / 59.0 IoU_cmpl, which sits outside that restriction.
The correction pass costs 107 ms, 3.23 FPS end to end. That is an incremental
pass, not a deployable rate.
Code, weights and the PS3 corpus are released. If something does not reproduce,
open an issue on the repo and I will look at it.
This is an automated message from the Librarian Bot. I found the following papers similar to this paper.
The following papers were recommended by the Semantic Scholar API
- FPSGen: Flexible Point Cloud Scene Generation with BEV-Supported Transport Flows (2026)
- BIMScript: Material-Aware Structured Scene Programs for BIM Ingestion (2026)
- Iterate or Widen? When Test-Time Refinement Helps LiDAR Scene Completion: A Controlled Study of Evidence Geometry, Training Coverage, and Compute (2026)
- Training-Free Open-Vocabulary 3D Point-Cloud Segmentation on the Generalized Few-Shot Benchmark (2026)
- GeoBridge: Decoupled Semantic Conditioning for Generative Image Geolocalization (2026)
- SparseOcc++: Geometry-Aware Sparse Latent Representation for Semantic Occupancy Prediction (2026)
- Towards Real-Time and Adaptable LiDAR Scene Completion (2026)
Please give a thumbs up to this comment if you found it helpful!
If you want recommendations for any Paper on Hugging Face checkout this Space
You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: @librarian-bot recommend
Get this paper in your agent:
hf papers read 2608.26737 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 1
Datasets citing this paper 2
Spaces citing this paper 0
No Space linking this paper