Pith. sign in

REVIEW 4 major objections 4 minor 29 references

Geometry-aware LoRA tuning makes a 2D instruction-following segmenter produce view-consistent masks that lift to faithful 3D reconstructions, adapting only 11.6M parameters.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Geometry-aware LoRA tuning of LISA with differentiable reprojection yields view-consistent masks that lift to better 3D reconstructions through frozen SAM-3D.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection A neat modular pipeline with a clean write-up, but the key ablation lacks a LoRA-only control, so the claimed boost from the geometric loss is not actually demonstrated. the 4 major comments →

arxiv 2512.01008 v2 pith:AYNCYYR5 submitted 2025-11-30 cs.CV

LISA-3D: Lifting Language-Image Segmentation to 3D via Multi-View Consistency

Zhongbin Guo , Jiahe Liu , Wenyu Gao , Yushan Li , Xiaomin He , Chengzhi Li , Ping Jian This is my paper
classification cs.CV
keywords LISA-3Dreferring expression segmentationmulti-view consistencyLoRAdifferentiable reprojection3D reconstructionGaussian splattingRGB-D
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 reading

The paper tries to prove that a 2D instruction-following segmenter can be made geometrically consistent across views with a lightweight adaptation, and that this consistency is the key to good language-driven 3D reconstruction. It retrofits LISA, a vision-language segmenter, with low-rank adapters and a differentiable reprojection loss that punishes mask disagreement between views, while leaving the 3D reconstructor SAM-3D frozen. If correct, this yields a modular pipeline where a user describes an object in natural language and gets an accurate Gaussian splat or mesh without any 3D-text supervision or retraining of the 3D module. The authors report gains of up to +15.6 points in 2D mIoU and large improvements in reconstruction F-score and Chamfer distance on ScanRefer and Nr3D, with only about 11.6M trainable parameters.

Core claim

On the paper's own terms, the central discovery is that geometry-aware LoRA tuning of a 2D segmenter, driven by a self-supervised reprojection-consistency loss, yields masks that stay consistent under viewpoint change, and that these masks, when formatted as RGBA prompts, let a frozen SAM-3D reconstructor produce significantly better 3D geometry. The authors show that the geometry loss alone (even with single-view inference) improves 2D mIoU from 10.2 to 17.6 on ScanRefer, and that adding multi-view training and inference raises it to 25.4, with F-score rising from 54.7 to 70.3 and Chamfer distance dropping from 12.4 to 7.9. The gains separate cleanly: training-time geometry awareness helps

What carries the argument

The load-bearing component is a differentiable warping operator that takes predicted logits from one view, unprojects them into 3D using depth and camera intrinsics, transforms them by the relative pose, and reprojects them into another view via bilinear sampling. The geometric loss compares the warped logits against the other view's prediction with stop-gradient, preventing the model from collapsing both views to the same trivial answer. LoRA adapters (rank 16) injected into every attention layer of LISA's vision and language branches carry the learned geometric awareness, keeping the trainable parameter count at about 11.6M. At deployment, a binary mask thresholded from the logits is conca

Load-bearing premise

The reprojection loss assumes that the depth maps and camera poses are accurate, so the warped masks are trustworthy targets.

What would settle it

Run training with artificially corrupted depth (e.g., +5% uniform noise) on ScanRefer and compare 2D mIoU; if the gain over the no-geometry-loss baseline disappears, the method's benefit comes from accurate RGB-D rather than the consistency principle.

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

If this is right

  • A 2D vision-language segmenter can acquire view consistency through a lightweight, self-supervised geometry loss, so mask-guided 3D reconstruction can reuse frozen 3D models.
  • The consistency objective acts as a regularizer that improves single-view masks, not only multi-view fusion.
  • Two training views capture most of the benefit; more views add marginal gains at higher cost.
  • The pipeline supports open-vocabulary referring expressions and zero-shot deployment on unseen categories, with only 11.6M trainable parameters.
  • Multi-view inference with logit fusion gives an additional boost beyond training-time geometry awareness.

Where Pith is reading between the lines

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

  • The same reprojection-consistency recipe could be applied to other 2D vision-language segmenters, suggesting that geometric self-supervision is a general way to lift 2D grounding models into 3D.
  • If depth and pose are noisy, an uncertainty-weighted warping or learned depth/pose refinement would likely be needed; the paper notes this as future work, implying the current gain is tied to clean RGB-D inputs.
  • The binary-mask interface outperforming soft probability maps hints that frozen 3D lifters are sensitive to prompt format; a learned interface from the segmenter's latent tokens could unlock further gains.
  • Extending the evaluation to outdoor or long-tail scenes would test whether the consistency loss transfers beyond indoor ScanNet-style data.
Share X Bluesky LinkedIn Reddit HN

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

4 major / 4 minor

Summary. LISA-3D adapts the instruction-following segmenter LISA using LoRA and a differentiable multi-view reprojection-consistency loss (L_geo), then feeds thresholded binary masks as RGBA prompts to a frozen SAM-3D reconstructor for object-centric 3D lifting. The method is trained on a filtered subset of ScanRefer RGB-D frames and evaluated on ScanRefer and Nr3D in terms of 2D mask mIoU, 3D F-score, and Chamfer distance. The paper claims up to +15.6 points mIoU over single-view baselines while updating only 11.6M parameters, and claims to separate geometry-aware training gains from optional multi-view inference gains.

Significance. If the attributed gains are genuine, the recipe is practically valuable: it retrofits a 2D vision-language model with geometric consistency without requiring 3D-text supervision and keeps the 3D reconstructor frozen. The writing is clear, the loss formulation in Eqs. (1)-(6) is standard and reproducible, and the qualitative demos suggest plausible qualitative improvement. However, the central causal claim—that L_geo, rather than ordinary supervised fine-tuning, drives the reported gains—is not established by the current ablation set, and the paper has no fine-tuned LISA baseline in the main table. These issues are load-bearing for the paper's stated contributions.

major comments (4)
  1. [Table 2 / §4.1 / Table 1] The ablation lacks a geometry-free LoRA-fine-tuned control. Row 1, '1 view, w/o L_geo', reports mIoU 10.2 and F-score 54.7, numerically identical to the untuned LISA baseline entries in Table 1 (mIoU 10.2; F-score 54.7 for LISA+SAM3D). Thus Row 1 is not a model trained with L_seg alone; it is the original LISA. Consequently, the jump 10.2→17.6 mIoU and 54.7→61.8 F-score attributed to 'introducing L_geo' could be entirely due to LoRA fine-tuning with BCE+Dice on the 27k filtered ScanRefer pairs. A control row with identical LoRA and data but λ=0 is required to support the claim that the geometric loss is responsible.
  2. [§4.1 / Table 2] The paper claims to 'separate geometry-aware training gains from optional multi-view inference gains,' but Row 2→Row 3 changes both the training view count (1→2) and the inference mode (single-view→multi-view fusion). Without configurations such as '2-view training + single-view inference' and '2-view training + multi-view inference' (both with L_geo), the +7.8 mIoU gain from Row 2 to Row 3 cannot be attributed to multi-view training, multi-view inference, or the combination. This separation is an explicit contribution and needs dedicated ablation rows.
  3. [§2.4 vs §3.2] The inference protocol is contradictory. §2.4 states 'During inference, only a single image I and text T are required; multi-view batches are used solely to train θ.' §3.2 states 'During inference, we sample two views per instruction, fuse predictions via our warping operator.' The fusion rule (reference frame, aggregation method, confidence weighting) is never specified. Because the main 'LISA-3D (Multi-view)' numbers in Table 1 use this multi-view inference, the results are not reproducible without a precise algorithm. Please provide the exact inference procedure and pseudocode for both single-view and multi-view deployment.
  4. [§3.1 / Table 1] The filtered training split is described only as 'approximately 27k high-quality utterance-view pairs where the target object is strictly visible and holds valid depth information.' The filtering criteria, number of removed pairs, and number of scenes excluded are not reported, making it impossible to assess potential selection bias. Relatedly, Table 1 contains no LISA fine-tuned baseline: the only 2D baseline is the untuned LISA (10.2 mIoU). The contribution bullet 'outperform LISA fine-tuning baselines' is therefore not demonstrated by the presented experiments. A LISA+LoRA model trained with L_seg only (equivalent to the missing control in the first major comment) should be included in Table 1.
minor comments (4)
  1. [Table 2] No error bars or significance tests are reported. The 3-view gain over 2 views (+0.8 mIoU, +0.8 F-score) is within plausible run-to-run variance, and the 'diminishing return' conclusion should be supported by repeated runs or statistical testing.
  2. [§4.2] The comparison of soft probability maps vs. discrete masks, and the embedding-level interface, is reported only qualitatively. A small table or quantitative statement would strengthen the practical guidance.
  3. [References] Reference [25] is listed as 'D Team' without author names; several references are arXiv preprints without version/date information (e.g., [24], [28]). Please standardize.
  4. [Figure 2] Figure 2 mentions 'disjoint RGBA prompts' but the figure does not show the RGBA input or the text prompts; adding these would help the reader connect the qualitative claim to the method.

Circularity Check

0 steps flagged

No circular reasoning; the reported gains are measured on held-out validation, not derived from the loss by construction. The Lgeo ablation lacks a geometry-free fine-tuned control, but that is an experimental-control weakness, not circularity.

full rationale

LISA-3D does not reduce its claimed result to its own inputs. The reprojection loss Lgeo in Eq. (6) is defined from depth maps, camera poses, and model predictions, and it is not a function of the reported validation metrics (2D mIoU, F-score, CD). The improvement attributed to geometry-aware tuning is evaluated on held-out ScanRefer/Nr3D validation splits, so the headline numbers are empirical outcomes rather than quantities forced by construction. The method builds on external prior work (LISA, SAM-3D, LoRA), and there are no author-overlapping self-citations, no imported uniqueness theorems, and no ansatz smuggled in via citation. Hyperparameters (lambda=0.4, LoRA rank r=16, threshold tau=0.5) are fixed choices and are not fitted to the evaluation metrics. The only substantive concern is attribution rather than circularity: Table 2's '1 view, w/o Lgeo' row exactly matches the untuned LISA baseline (10.2 mIoU, 54.7 F-score), so the gain to 17.6/61.8 with Lgeo also includes LoRA fine-tuning and supervised Lseg on the filtered training pairs. Thus the statement in Sec. 4.1 that 'introducing Lgeo yields substantial improvements' is not cleanly isolated, but the improvement is not equivalent by definition to the loss itself. This is a missing-control weakness, not a circular derivation, so the score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The method introduces no new entities. Its main free parameters are the loss weight and LoRA hyperparameters; the use of ground-truth masks is an external supervision source that the abstract underplays.

free parameters (4)
  • λ (geometric loss weight) = 0.4
    Hand-chosen weight balancing L_seg and L_geo; no sensitivity analysis reported (Sec. 3.2, Eq. 7).
  • LoRA rank r and α = r=16, α=32
    PEFT hyperparameters chosen without ablations (Sec. 3.2).
  • Mask threshold τ = 0.5
    Binarization threshold for RGBA prompt; fixed in all experiments (Sec. 2.3).
  • Training view count = 2
    Selected by ablation after observing diminishing returns with 3 views (Table 2).
axioms (4)
  • standard math Camera projection and bilinear warping equations (Eq. 1-4) are correct and differentiable.
    Standard pinhole camera model and grid_sample warping; used to build the reprojection loss.
  • domain assumption RGB-D sequences provide accurate depth maps and camera poses.
    The geometry loss relies on depth and extrinsics being reliable; the paper notes in Limitations that estimation errors could destabilize the loss.
  • domain assumption Ground-truth masks derived from ScanRefer/Nr3D 3D annotations correctly supervise 2D segmentation.
    L_seg uses projected ground-truth masks (Eq. 5), so the method is not fully self-supervised despite the abstract's wording.
  • ad hoc to paper Stop-gradient in Eq. (6) prevents trivial collapse without biasing the training signal.
    A heuristic to avoid both predictions collapsing to zero; no formal analysis is provided.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of LISA-3D: Lifting Language-Image Segmentation to 3D via Multi-View Consistency." pith.science (2026). https://pith.science/paper/AYNCYYR5

@misc{pith2026251201008,
  author       = {Pith},
  title        = {Pith review of: LISA-3D: Lifting Language-Image Segmentation to 3D via Multi-View Consistency},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AYNCYYR5}},
  note         = {Machine review of arXiv:2512.01008}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Text-driven 3D reconstruction requires masks that understand free-form instructions and remain stable under viewpoint changes. We present LISA-3D, a two-stage framework that adapts the instruction-following segmenter LISA with geometry-aware Low-Rank Adaptation (LoRA) layers while keeping the SAM-3D reconstructor frozen. During training, paired RGB-D frames and camera poses define a differentiable reprojection loss that enforces cross-view agreement without additional 3D-text annotations. At deployment, the adapted segmenter can produce an RGBA prompt for SAM-3D from one RGB image; when registered RGB-D views are available, optional logit fusion further improves the prompt. On ScanRefer and Nr3D, geometry-aware tuning improves both 2D masks and lifted 3D reconstructions while updating only 11.6M parameters. Our results separate geometry-aware training gains from optional multi-view inference gains, providing a modular route from language grounding to object-centric 3D reconstruction.

Figures

Figures reproduced from arXiv: 2512.01008 by Chengzhi Li, Jiahe Liu, Ping Jian, Wenyu Gao, Xiaomin He, Yushan Li, Zhongbin Guo.

Figure 1
Figure 1. Figure 1: Overview of LISA-3D. Our framework operates in two stages. First, geometry-aware LISA receives multi-view RGB-D pairs [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative demos. From left to right: input image, LISA-3D mask, SAM-3D reconstruction (Gaussian splat) and textured mesh. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] 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

29 extracted references · 4 linked inside Pith

  1. [1]

    ReferIt3D: Neural Listen- ers for Fine-Grained 3D Object Identification in Real-World Scenes

    Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. ReferIt3D: Neural Listen- ers for Fine-Grained 3D Object Identification in Real-World Scenes. InComputer Vision – ECCV 2020, pages 422–440. Springer International Publishing, Cham, 2020. 4, 7

    2020

  2. [2]

    Qwen2.5-VL Technical Report, 2025

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhao- hai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-VL Technical Report, 2025. 7

    2025

  3. [3]

    From thousands to billions: 3d visual language ground- ing via render-supervised distillation from 2d VLMs

    Ang Cao, Sergio Arnaud, Oleksandr Maksymets, Jianing Yang, Ayush Jain, Ada Martin, Vincent-Pierre Berges, Paul McVay, Ruslan Partsey, Aravind Rajeswaran, Franziska Meier, Justin Johnson, Jeong Joon Park, and Alexander Sax. From thousands to billions: 3d visual language ground- ing via render-supervised distillation from 2d VLMs. In Forty-second Internatio...

    2025

  4. [4]

    SAM 3: Segment Anything with Concepts, 2025

    Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoub- hik Debnath, Ronghang Hu, Didac Suris, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, Andrew Huang, Jie Lei, Tengyu Ma, Baishan Guo, Arpit Kalla, Markus Marks, Joseph Greer, Meng Wang, Peize Sun, Roman R¨adle, Triantafyllos Afouras, Effrosyni Mavroudi, Kather- ine Xu, Tsung-Han Wu, Yu Zhou, Lil...

    2025

  5. [5]

    Scanrefer: 3d object localization in rgb-d scans using natural language

    Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. InEuropean conference on computer vision, pages 202–221. Springer, 2020. 4, 7

    2020

  6. [6]

    Sdfusion: Multimodal 3d shape completion, reconstruction, and generation

    Yen-Chi Cheng, Hsin-Ying Lee, Sergey Tulyakov, Alexan- der G Schwing, and Liang-Yan Gui. Sdfusion: Multimodal 3d shape completion, reconstruction, and generation. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4456–4465, 2023. 7

    2023

  7. [7]

    Lam3d: Large image-point clouds align- ment model for 3d reconstruction from single image.Ad- vances in Neural Information Processing Systems, 37:4454– 4480, 2024

    Ruikai Cui, Xibin Song, Weixuan Sun, Senbo Wang, Weizhe Liu, Shenzhou Chen, Taizhang Shang, Yang Li, Nick Barnes, Hongdong Li, et al. Lam3d: Large image-point clouds align- ment model for 3d reconstruction from single image.Ad- vances in Neural Information Processing Systems, 37:4454– 4480, 2024. 7

    2024

  8. [8]

    Scannet: Richly-annotated 3d reconstructions of indoor scenes

    Angela Dai, Angel X Chang, Manolis Savva, Maciej Hal- ber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828–5839, 2017. 4, 7

    2017

  9. [9]

    Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022. 2

    2022

  10. [10]

    SceneVerse: Scaling 3D Vision-Language Learning for Grounded Scene Understanding

    Baoxiong Jia, Yixin Chen, Huangyue Yu, Yan Wang, Xuesong Niu, Tengyu Liu, Qing Li, and Siyuan Huang. SceneVerse: Scaling 3D Vision-Language Learning for Grounded Scene Understanding. InComputer Vision – ECCV 2024, pages 289–310. Springer Nature Switzerland, Cham, 2025. 1

    2024

  11. [11]

    3d gaussian splatting for real-time radiance field rendering.ACM Trans

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Trans. Graph., 42(4):139–1,

  12. [12]

    Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross Girshick

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C. Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross Girshick. Segment Anything, 2023. 1, 7

    2023

  13. [13]

    LISA: Reasoning Segmen- tation via Large Language Model, 2024

    Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. LISA: Reasoning Segmen- tation via Large Language Model, 2024. 1, 4, 7

    2024

  14. [14]

    Part123: part-aware 3d reconstruction from a single-view image

    Anran Liu, Cheng Lin, Yuan Liu, Xiaoxiao Long, Zhiyang Dou, Hao-Xiang Guo, Ping Luo, and Wenping Wang. Part123: part-aware 3d reconstruction from a single-view image. InACM SIGGRAPH 2024 Conference Papers, pages 1–12, 2024. 7

    2024

  15. [15]

    Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023. 7

    2023

  16. [16]

    Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017. 4

    Pith/arXiv arXiv 2017

  17. [17]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021. 7 8

    2021

  18. [18]

    SAM 2: Segment Anything in Images and Videos,

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman R¨adle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junt- ing Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao- Yuan Wu, Ross Girshick, Piotr Doll´ar, and Christoph Feicht- enhofer. SAM 2: Segment Anything in Images and Videos,

  19. [19]

    A survey of language-grounded mul- timodal 3d scene understanding.Knowledge-Based Systems, 321:113650, 2025

    Ruilong Ren, Xinyu Zhao, Weichen Xu, Jian Cao, Xinxin Xu, and Xing Zhang. A survey of language-grounded mul- timodal 3d scene understanding.Knowledge-Based Systems, 321:113650, 2025. 1

    2025

  20. [20]

    Grounded sam: Assembling open-world models for diverse visual tasks.arXiv preprint arXiv:2401.14159,

    Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks.arXiv preprint arXiv:2401.14159,

    Pith/arXiv arXiv

  21. [21]

    Anything- 3d: Towards single-view anything reconstruction in the wild

    Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Anything- 3d: Towards single-view anything reconstruction in the wild. arXiv preprint arXiv:2304.10261, 2023. 7

    Pith/arXiv arXiv 2023

  22. [22]

    Point-gnn: Graph neural net- work for 3d object detection in a point cloud

    Weijing Shi and Raj Rajkumar. Point-gnn: Graph neural net- work for 3d object detection in a point cloud. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1711–1719, 2020. 7

    2020

  23. [23]

    Splat-MOVER: Multi-Stage, Open- V ocabulary Robotic Manipulation via Editable Gaussian Splatting

    Ola Shorinwa, Johnathan Tucker, Aliyah Smith, Aiden Swann, Timothy Chen, Roya Firoozi, Monroe Kennedy Iii, and Mac Schwager. Splat-MOVER: Multi-Stage, Open- V ocabulary Robotic Manipulation via Editable Gaussian Splatting. 1

  24. [24]

    Natural Language Guided Goals for Robotic Manipulation

    Andrew Nichols Crawford Taylor. Natural Language Guided Goals for Robotic Manipulation. 1

  25. [25]

    SAM 3D: 3Dfy Anything in Images

    D Team, Xingyu Chen, Fu-Jen Chu, Pierre Gleize, Kevin J Liang, Alexander Sax, Hao Tang, Weiyao Wang, Michelle Guo, Thibaut Hardin, Xiang Li, Aohan Lin, Jiawei Liu, Ziqi Ma, Anushka Sagar, Bowen Song, Xiaodong Wang, Jianing Yang, Bowen Zhang, Piotr Doll´ar, Georgia Gkioxari, and Ji- tendra Malik. SAM 3D: 3Dfy Anything in Images. 1, 4, 7

  26. [26]

    Sa2V A: Marrying SAM2 with LLaV A for Dense Grounded Understanding of Images and Videos, 2025

    Haobo Yuan, Xiangtai Li, Tao Zhang, Yueyi Sun, Zilong Huang, Shilin Xu, Shunping Ji, Yunhai Tong, Lu Qi, Jiashi Feng, and Ming-Hsuan Yang. Sa2V A: Marrying SAM2 with LLaV A for Dense Grounded Understanding of Images and Videos, 2025. 1

    2025

  27. [27]

    Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual refer- ring

    Zhihao Yuan, Xu Yan, Yinghong Liao, Ruimao Zhang, Sheng Wang, Zhen Li, and Shuguang Cui. Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual refer- ring. InProceedings of the IEEE/CVF International Confer- ence on Computer Vision, pages 1791–1800, 2021. 7

    2021

  28. [28]

    3D- VLA: A 3D Vision-Language-Action Generative World Model

    Haoyu Zhen, Xiaowen Qiu, Peihao Chen, Jincheng Yang, Xin Yan, Yilun Du, Yining Hong, and Chuang Gan. 3D- VLA: A 3D Vision-Language-Action Generative World Model. https://arxiv.org/abs/2403.09631v1, 2024. 1

    Pith/arXiv arXiv 2024

  29. [29]

    EditRoom: LLM-parameterized Graph Diffusion for Composable 3D Room Layout Editing,

    Kaizhi Zheng, Xiaotong Chen, Xuehai He, Jing Gu, Lin- jie Li, Zhengyuan Yang, Kevin Lin, Jianfeng Wang, Lijuan Wang, and Xin Eric Wang. EditRoom: LLM-parameterized Graph Diffusion for Composable 3D Room Layout Editing,

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.