Pith. sign in

REVIEW 3 major objections 1 minor 44 references

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.3

Reshaping 3D semantic occupancy grids into multi-channel BEV images lets Stable Diffusion generate and edit scenes without 3D-specific networks or retraining.

2026-06-27 17:06 UTC pith:W3WKFGMO

load-bearing objection EditSSC shows a straightforward way to reuse Stable Diffusion for 3D semantic occupancy via BEV reshaping and codebook editing, but the abstract gives no numbers so the outperformance claim stays unverified. the 3 major comments →

arxiv 2606.09273 v1 pith:W3WKFGMO submitted 2026-06-08 cs.CV

EditSSC: Toward Editable Semantic Occupancy Scenes with Unconditional Diffusion Models

Fatima Balde , Raoul de Charette , Alexandre Boulch This is my paper
classification cs.CV
keywords semantic occupancylatent diffusionBEV representationscene editingSemanticKITTIunconditional generationinpainting
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that standard 2D latent diffusion models can handle 3D semantic scene tasks for autonomous driving by first converting the 3D grids into 2D bird's eye view images. This conversion allows the use of an off-the-shelf quantized autoencoder and UNet, with diffusion performed in the latent space after quantization. The approach yields unconditional generation that exceeds 3D baselines on SemanticKITTI while also enabling sketch-guided generation, inpainting, and outpainting through class-to-code mappings in the codebook. A sympathetic reader would care because it removes the need for custom 3D architectures and supports editing out of the box.

Core claim

EditSSC shows that 3D semantic occupancy scenes can be generated unconditionally and edited via sketch guidance, inpainting, and outpainting by reshaping the grids into multi-channel BEV images, running them through Stable Diffusion's existing autoencoder and UNet with diffusion on quantized latents, and exploiting class-to-code correspondences, all without retraining or 3D modifications.

What carries the argument

Reshaping 3D semantic occupancy grids into multi-channel BEV images passed through Stable Diffusion's quantized autoencoder and UNet for latent diffusion, which supports training-free editing by using class-to-code correspondences in the codebook.

Load-bearing premise

Converting 3D occupancy grids to 2D BEV images and processing them with a 2D diffusion model retains enough 3D spatial structure for generation and editing to stay useful.

What would settle it

On the SemanticKITTI validation set, EditSSC produces lower mIoU or higher FID scores than the 3D baselines it claims to beat, or edited outputs show clear loss of vertical 3D consistency such as incorrect object heights or stacking.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Unconditional generation on SemanticKITTI outperforms existing 3D-specific baselines.
  • Sketch-guided generation, inpainting, and outpainting become possible without any retraining.
  • Only minimal changes are needed to adapt an existing 2D diffusion pipeline for 3D semantic occupancy work.
  • Well-established 2D architectures can be repurposed directly for 3D scene generation and editing tasks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar BEV reshaping might let other pre-trained 2D models handle additional 3D perception tasks if height information is not critical.
  • The method could lower barriers for creating editable driving scene simulators by avoiding custom 3D training.
  • If vertical structure is lost in the projection, extensions that add explicit height channels or multi-view inputs might restore accuracy.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 1 minor

Summary. The paper introduces EditSSC, a method for unconditional 3D semantic occupancy scene generation that reshapes voxel grids into multi-channel BEV images, encodes them via Stable Diffusion's pre-trained VQVAE and UNet with minimal modifications, and performs diffusion in latent space. It claims superior performance over 3D-specific baselines on SemanticKITTI for unconditional generation and supports training-free sketch-guided generation, inpainting, and outpainting by leveraging class-to-code correspondences in the codebook.

Significance. If the central claims hold, the work would be significant for showing that established 2D latent diffusion pipelines can be repurposed for 3D semantic tasks in autonomous driving, reducing the need for custom 3D architectures while adding practical editing functionality. The training-free editing via codebook correspondences is a clear strength that could enable flexible scene manipulation without retraining.

major comments (3)
  1. [Approach paragraph] Approach paragraph (and abstract): The pipeline encodes multi-channel BEV images derived from semantic occupancy grids using Stable Diffusion's off-the-shelf quantized autoencoder without channel-wise adaptation, fine-tuning, or a semantic-specific codebook. This step is load-bearing for both the unconditional generation quality (needed to outperform 3D baselines) and the editing operations (which rely on faithful class-to-code mappings); if discrete labels are not preserved, the reported advantages become unreliable.
  2. [Abstract] Abstract and results (implied): The claim that EditSSC 'outperforms existing 3D-specific baselines on unconditional generation' on SemanticKITTI is stated without any quantitative metrics, error bars, dataset splits, ablation details, or comparison tables. These details are required to evaluate whether the BEV + pre-trained VQVAE approach actually supports the performance claim.
  3. [Approach paragraph] Approach paragraph: The assumption that reshaping 3D occupancy grids into multi-channel BEV images and passing them through the RGB-trained VQVAE/UNet 'preserves enough 3D spatial structure' for useful generation and editing is not supported by any analysis or validation experiment in the provided description, yet it underpins the entire method.
minor comments (1)
  1. [Abstract] The abstract would benefit from a brief statement of the specific metrics used to claim outperformance (e.g., mIoU, FID) even if full tables appear later.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below with clarifications and note planned revisions to the manuscript.

read point-by-point responses

  1. Referee: [Approach paragraph] Approach paragraph (and abstract): The pipeline encodes multi-channel BEV images derived from semantic occupancy grids using Stable Diffusion's off-the-shelf quantized autoencoder without channel-wise adaptation, fine-tuning, or a semantic-specific codebook. This step is load-bearing for both the unconditional generation quality (needed to outperform 3D baselines) and the editing operations (which rely on faithful class-to-code mappings); if discrete labels are not preserved, the reported advantages become unreliable.

    Authors: We agree that faithful preservation of semantic classes through the off-the-shelf VQVAE is essential. The manuscript demonstrates this indirectly via strong generation and editing results that rely on accurate class-to-code mappings. To directly validate the point, we will add a reconstruction fidelity analysis for semantic labels in the revised version. revision: yes

  2. Referee: [Abstract] Abstract and results (implied): The claim that EditSSC 'outperforms existing 3D-specific baselines on unconditional generation' on SemanticKITTI is stated without any quantitative metrics, error bars, dataset splits, ablation details, or comparison tables. These details are required to evaluate whether the BEV + pre-trained VQVAE approach actually supports the performance claim.

    Authors: The experiments section of the full manuscript includes quantitative tables, metrics, dataset details, and ablations on SemanticKITTI. To strengthen the abstract, we will add key performance numbers and a reference to the comparison table. revision: partial

  3. Referee: [Approach paragraph] Approach paragraph: The assumption that reshaping 3D occupancy grids into multi-channel BEV images and passing them through the RGB-trained VQVAE/UNet 'preserves enough 3D spatial structure' for useful generation and editing is not supported by any analysis or validation experiment in the provided description, yet it underpins the entire method.

    Authors: Empirical success on unconditional generation (outperforming 3D baselines) and training-free editing provides supporting evidence. We will add an explicit validation experiment or spatial consistency analysis in the revised manuscript to directly address this assumption. revision: yes

Circularity Check

0 steps flagged

No circularity; reuses external pre-trained Stable Diffusion components with empirical claims

full rationale

The paper's core approach reshapes 3D semantic occupancy grids into multi-channel BEV images and applies Stable Diffusion's off-the-shelf quantized autoencoder and UNet with minimal modifications, performing diffusion on the resulting latents. This reuses an external pre-trained model rather than deriving any quantity from fitted parameters internal to the work. No equations, self-definitional mappings, or predictions that reduce to inputs by construction appear in the abstract or described method. Performance claims (outperforming 3D baselines on SemanticKITTI) and editing capabilities (sketch-guided generation via class-to-code correspondences) are presented as empirical outcomes, not as derivations forced by self-citation chains or ansatzes. The derivation chain is self-contained against external benchmarks and does not exhibit any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are stated. The approach implicitly assumes that the pre-trained Stable Diffusion components transfer without modification to the BEV domain.

pith-pipeline@v0.9.1-grok · 5703 in / 1185 out tokens · 20366 ms · 2026-06-27T17:06:15.022675+00:00 · methodology

0 comments
read the original abstract

3D semantic scene generation is crucial for autonomous driving applications, yet most methods rely on complex 3D-specific architectures such as triplane encoders and adapted diffusion networks, limiting both their simplicity and their editing capabilities. We propose EditSSC, an editing-ready method for 3D semantic scene generation using 2D Bird's Eye View (BEV) representations and off-the-shelf latent diffusion network. Our approach reshapes 3D semantic occupancy grids into multi-channel BEV images and leverages the quantized autoencoder and UNet from Stable Diffusion with minimal modifications. We perform diffusion on the latents after quantization, which enables training-free editing capabilities. By exploiting class-to-code correspondences in the codebook, our method supports sketch-guided generation, inpainting, and outpainting without any retraining. On SemanticKITTI, EditSSC outperforms existing 3D-specific baselines on unconditional generation, demonstrating that well-established 2D architectures can be effectively repurposed for 3D scene generation and editing.

Figures

Figures reproduced from arXiv: 2606.09273 by Alexandre Boulch, Fatima Balde, Raoul de Charette.

Figure 1
Figure 1. Figure 1: EditSSC capabilities. Our scene generation method relies on a latent diffusion model with carefully designed components to enable training-free editing capability. While unconditional scene generation (left) can produce multiple diverse samples, all generated scenes faithfully follow the training distribution. EditSSC further enables editing via sketch guidance (using a user-provided layout), inpainting, a… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of EditSSC. The training consists of two stages. In the first stage (Sec. 4.1), a 3D semantic occupancy scene is passed through a VQ-VAE that compresses it into discrete latent codes. In the second stage (Sec. 4.2), a lightweight U-Net performs diffusion on the quantized latents to generate new scenes. The discrete codebook further enables training-free editing via class-to-code correspondences (S… view at source ↗
Figure 3
Figure 3. Figure 3: Unconditional generation. Our method generates plau￾sible scenes (bottom) which follow the class distribution and gen￾eral structure of the training set (top) [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: LiDAR-conditioned generation. Given a LiDAR scan (left), our model generates a semantic occupancy scene (mid￾dle) which resembles the ground truth (right). 5.2. Unconditional generation. For unconditional generation, we sample a random BEV latent zT ∼ N (0, I) and iteratively denoise it using the trained UNet following the DDPM [10] reverse process. The resulting latent z0 is then passed through the VQ-VAE… view at source ↗
Figure 5
Figure 5. Figure 5: Training-free sketch-guided generation. Given a user-drawn BEV layout, our method generates diverse and coherent 3D semantic scenes without any retraining. Each layout is shown with two independently generated scenes, demonstrating both fidelity to the user-specified structure and diversity in the completions. Sketch-guided generation. We illustrate the sketch￾guided generation capability in [PITH_FULL_IM… view at source ↗
Figure 6
Figure 6. Figure 6: Scene inpainting and outpainting. Inpainting (top): given a scene with a masked region, our model generates coherent content to fill the missing area while preserving the known regions. Outpainting (bottom): given half of an existing scene, our model extends it by generating the missing half, producing a spatially coherent continuation. Autoencoder Diffusion Codes Dim. Codebook utilization ↑ IoU↑ mIoU↑ FID… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

44 extracted references · 3 canonical work pages · 2 internal anchors

  1. [1]

    Cc3d: Layout-conditioned generation of compositional 3d scenes

    Sherwin Bahmani, Jeong Joon Park, Despoina Paschalidou, Xingguang Yan, Gordon Wetzstein, Leonidas Guibas, and Andrea Tagliasacchi. Cc3d: Layout-conditioned generation of compositional 3d scenes. InICCV, 2023. 2

    2023

  2. [2]

    Se- mantickitti: A dataset for semantic scene understanding of lidar sequences

    Jens Behley, Martin Garbade, Andres Milioto, Jan Quen- zel, Sven Behnke, Cyrill Stachniss, and Jurgen Gall. Se- mantickitti: A dataset for semantic scene understanding of lidar sequences. InICCV, 2019. 2, 5

    2019

  3. [3]

    The lov ´asz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks

    Maxim Berman, Amal Rannen Triki, and Matthew B Blaschko. The lov ´asz-softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks. InCVPR, 2018. 4 8

    2018

  4. [4]

    Demystifying mmd gans

    Mikołaj Bi´nkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. InICLR, 2018. 5

    2018

  5. [5]

    Diffssc: Semantic lidar scan completion using denoising diffusion probabilistic models

    Helin Cao and Sven Behnke. Diffssc: Semantic lidar scan completion using denoising diffusion probabilistic models. InIROS, 2025. 2, 6

    2025

  6. [6]

    Scaling recti- fied flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas M ¨uller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling recti- fied flow transformers for high-resolution image synthesis. InICML, 2024. 1

    2024

  7. [7]

    Two stream 3d semantic scene completion

    Martin Garbade, Yueh-Tung Chen, Johann Sawatzky, and Juergen Gall. Two stream 3d semantic scene completion. InCVPR-W, 2019. 6

    2019

  8. [8]

    Gans trained by a two time-scale update rule converge to a local nash equilib- rium.NeurIPS, 30, 2017

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- rium.NeurIPS, 30, 2017. 5

    2017

  9. [9]

    Classifier-Free Diffusion Guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  10. [10]

    Denoising diffu- sion probabilistic models.NeurIPS, 33, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models.NeurIPS, 33, 2020. 2, 4, 6

    2020

  11. [11]

    Diffindscene: Diffusion-based high-quality 3d indoor scene generation

    Xiaoliang Ju, Zhaoyang Huang, Yijin Li, Guofeng Zhang, Yu Qiao, and Hongsheng Li. Diffindscene: Diffusion-based high-quality 3d indoor scene generation. InCVPR, 2024. 2

    2024

  12. [12]

    Shap-E: Generating Conditional 3D Implicit Functions

    Heewoo Jun and Alex Nichol. Shap-e: Generat- ing conditional 3d implicit functions.arXiv preprint arXiv:2305.02463, 2023. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  13. [13]

    Neuralfield-ldm: Scene gener- ation with hierarchical latent diffusion models

    Seung Wook Kim, Bradley Brown, Kangxue Yin, Karsten Kreis, Katja Schwarz, Daiqing Li, Robin Rombach, Antonio Torralba, and Sanja Fidler. Neuralfield-ldm: Scene gener- ation with hierarchical latent diffusion models. InCVPR,

  14. [14]

    Eq-vae: Equivariance regularized latent space for improved generative image modeling

    Theodoros Kouzelis, Ioannis Kakogeorgiou, Spyros Gidaris, and Nikos Komodakis. Eq-vae: Equivariance regularized latent space for improved generative image modeling. In ICML, 2025. 2

    2025

  15. [15]

    Diffusion Probabilistic Models for Scene-Scale 3D Categorical Data

    Jumin Lee, Woobin Im, Sebin Lee, and Sung-Eui Yoon. Dif- fusion probabilistic models for scene-scale 3d categorical data.arXiv preprint arXiv:2301.00527, 2023. 2, 3

    work page Pith review arXiv 2023

  16. [16]

    Semcity: Semantic scene gener- ation with triplane diffusion

    Jumin Lee, Sebin Lee, Changho Jo, Woobin Im, Juhyeong Seon, and Sung-Eui Yoon. Semcity: Semantic scene gener- ation with triplane diffusion. InCVPR, 2024. 1, 2, 3, 5

    2024

  17. [17]

    Latent diffusion models with masked autoencoders

    Junho Lee, Jeongwoo Shin, Hyungwook Choi, and Joonseok Lee. Latent diffusion models with masked autoencoders. In ICCV, 2025. 2, 3

    2025

  18. [18]

    Diffusion- sdf: Text-to-shape via voxelized diffusion

    Muheng Li, Yueqi Duan, Jie Zhou, and Jiwen Lu. Diffusion- sdf: Text-to-shape via voxelized diffusion. InCVPR, 2023. 2

    2023

  19. [19]

    Diffusiondrive: Truncated diffusion model for end-to-end autonomous driving

    Bencheng Liao, Shaoyu Chen, Haoran Yin, Bo Jiang, Cheng Wang, Sixu Yan, Xinbang Zhang, Xiangyu Li, Ying Zhang, Qian Zhang, et al. Diffusiondrive: Truncated diffusion model for end-to-end autonomous driving. InCVPR, 2025. 2

    2025

  20. [20]

    A dual-path network for semantic scene com- pletion of single-frame lidar point clouds.Int

    Wei Liu, Ziwen Kang, Yongtao Yu, Zheng Gong, Yuchao Zheng, Xiaohui Huang, Haiyan Guan, Lingfei Ma, and De- dong Zhang. A dual-path network for semantic scene com- pletion of single-frame lidar point clouds.Int. J. Appl. Earth Obs., 2026. 6

    2026

  21. [21]

    Pyramid diffusion for fine 3d large scene generation

    Yuheng Liu, Xinke Li, Xueting Li, Lu Qi, Chongshou Li, and Ming-Hsuan Yang. Pyramid diffusion for fine 3d large scene generation. InECCV. Springer, 2024. 2

    2024

  22. [22]

    Meshdiffu- sion: Score-based generative 3d mesh modeling

    Zhen Liu, Yao Feng, Michael J Black, Derek Nowrouzezahrai, Liam Paull, and Weiyang Liu. Meshdiffu- sion: Score-based generative 3d mesh modeling. InICLR,

  23. [23]

    Matswap: Light-aware material transfers in images

    Ivan Lopes, Valentin Deschaintre, Yannick Hold-Geoffroy, and Raoul de Charette. Matswap: Light-aware material transfers in images. InComputer Graphics Forum. Wiley Online Library, 2025. 2

    2025

  24. [24]

    Repaint: Inpainting using denoising diffusion probabilistic models

    Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. InCVPR,

  25. [25]

    Diffusion probabilistic models for 3d point cloud generation

    Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. InCVPR, 2021. 2

    2021

  26. [26]

    Shapeshifter: 3d variations using mul- tiscale and sparse point-voxel diffusion

    Nissim Maruani, Wang Yifan, Matthew Fisher, Pierre Alliez, and Mathieu Desbrun. Shapeshifter: 3d variations using mul- tiscale and sparse point-voxel diffusion. InCVPR, 2025. 2

    2025

  27. [27]

    Lt3sd: Latent trees for 3d scene diffusion

    Quan Meng, Lei Li, Matthias Nießner, and Angela Dai. Lt3sd: Latent trees for 3d scene diffusion. InCVPR, 2025. 2

    2025

  28. [28]

    Maxime Oquab, Timoth ´ee Darcet, Theo Moutakanni, Huy V . V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Rus- sell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang- Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nico- las Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patri...

    2023

  29. [29]

    Xcube: Large-scale 3d generative modeling using sparse voxel hierarchies

    Xuanchi Ren, Jiahui Huang, Xiaohui Zeng, Ken Museth, Sanja Fidler, and Francis Williams. Xcube: Large-scale 3d generative modeling using sparse voxel hierarchies. In CVPR, 2024. 2

    2024

  30. [30]

    LMSCNet: Lightweight multiscale 3D semantic completion

    Luis Roldao, Raoul De Charette, and Anne Verroust- Blondet. LMSCNet: Lightweight multiscale 3D semantic completion. In3DV, 2020. 6

    2020

  31. [31]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. InCVPR, 2022. 1, 2, 3, 4

    2022

  32. [32]

    Palette: Image-to-image diffusion models

    Chitwan Saharia, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. InACM SIGGRAPH 2022 conference proceedings, 2022. 2

    2022

  33. [33]

    Photorealistic text-to-image diffusion models with deep language understanding.NeurIPS, 35, 2022

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding.NeurIPS, 35, 2022. 2

    2022

  34. [34]

    Diffusion-based signed distance fields for 3d shape gener- ation

    Jaehyeok Shim, Changwoo Kang, and Kyungdon Joo. Diffusion-based signed distance fields for 3d shape gener- ation. InCVPR, 2023. 2

    2023

  35. [35]

    3d neural field generation using triplane diffusion

    J Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3d neural field generation using triplane diffusion. InCVPR, 2023. 1, 2 9

    2023

  36. [36]

    Diffuscene: Denoising diffu- sion models for generative indoor scene synthesis

    Jiapeng Tang, Yinyu Nie, Lev Markhasin, Angela Dai, Justus Thies, and Matthias Nießner. Diffuscene: Denoising diffu- sion models for generative indoor scene synthesis. InCVPR,

  37. [37]

    Lion: Latent point dif- fusion models for 3d shape generation.NeurIPS, 35, 2022

    Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point dif- fusion models for 3d shape generation.NeurIPS, 35, 2022. 1, 2

    2022

  38. [38]

    Neural discrete representation learning.NeurIPS, 30, 2017

    Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning.NeurIPS, 30, 2017. 4

    2017

  39. [39]

    Blockfusion: Expandable 3d scene gen- eration using latent tri-plane extrapolation.ACM Transac- tions on Graphics (ToG), 43(4), 2024

    Zhennan Wu, Yang Li, Han Yan, Taizhang Shang, Weixuan Sun, Senbo Wang, Ruikai Cui, Weizhe Liu, Hiroyuki Sato, Hongdong Li, et al. Blockfusion: Expandable 3d scene gen- eration using latent tri-plane extrapolation.ACM Transac- tions on Graphics (ToG), 43(4), 2024. 1, 2

    2024

  40. [40]

    Flowssc: Universal generative monocular semantic scene completion via one-step latent diffusion.Robotics and Automation Letters, 2026

    Zichen Xi, Hao-Xiang Chen, Nan Xue, Hongyu Yan, Qi- Yuan Feng, Levent Burak Kara, Joaquim Jorge, and Qun- Ce Xu. Flowssc: Universal generative monocular semantic scene completion via one-step latent diffusion.Robotics and Automation Letters, 2026. 2, 3

    2026

  41. [41]

    Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion

    Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. InAAAI, 2021. 6

    2021

  42. [42]

    Representation alignment for generation: Training diffusion transformers is easier than you think

    Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think. InICLR, 2024. 2

    2024

  43. [43]

    Sseditor: Controllable mask-to-scene generation with diffusion model.Knowledge-Based Systems, 2026

    Haowen Zheng, Jiahao Pang, Zhiqiang Pu, and Yanyan Liang. Sseditor: Controllable mask-to-scene generation with diffusion model.Knowledge-Based Systems, 2026. 2, 6

    2026

  44. [44]

    3d shape genera- tion and completion through point-voxel diffusion

    Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape genera- tion and completion through point-voxel diffusion. InICCV,