Pith. sign in

REVIEW 3 major objections 4 minor 5 cited by

A single feedforward network with test-time training lets users edit 3D Gaussian splats from ordinary 2D strokes, at interactive speeds and without destroying the original asset.

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 ↗

T0 review · deepseek-v4-flash

2026-08-03 18:21 UTC pith:WX6JVTFU

load-bearing objection SplatPainter's core mechanism is a real step forward for interactive 3DGS editing, but the paper's own Table 1 contradicts its 'best quality' claim—needs a recalibration, not a rejection. the 3 major comments →

arxiv 2512.05354 v2 pith:WX6JVTFU submitted 2025-12-05 cs.CV cs.GR

SplatPainter: Interactive Authoring of 3D Gaussians from 2D Edits via Test-Time Training

Yang Zheng , Hao Tan , Kai Zhang , Peng Wang , Leonidas Guibas , Gordon Wetzstein , Wang Yifan This is my paper
classification cs.CV cs.GR
keywords 3D Gaussian splattingtest-time trainingfeedforward editinginteractive 3D authoringlocal detail refinementglobal relightingvoxel latent representation2D-edit-to-3D
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 sets out to close a gap in the 3D content pipeline: after a Gaussian-splat asset is generated, artists lack a fast, non-destructive way to refine it with familiar 2D tools. SplatPainter claims that a single feedforward architecture can take a 2D edit—a zoomed-in view for detail refinement, a graffiti stroke, or a relit view—and directly predict updates to the Gaussian attributes, without per-scene optimization or regeneration from diffusion. The key move is to convert the asset once into a compact, feature-rich voxel-latent representation, then use test-time training to adapt a small set of 'fast weights' to each user edit and apply the adapted transform to the latent. This yields a continuous, iterative workflow: each edit takes roughly 0.3 seconds for local changes and 0.06 seconds for global relighting, and the original unedited regions are preserved. If the claim holds, it makes 3D authoring feel like editing an image.

Core claim

The central claim is that updating a 3D Gaussian asset from user-supplied 2D images can be cast as a feedforward prediction problem, solved by a state-aware network, rather than an optimization or generative-regeneration problem. The pipeline renders the input splat from canonical views, reconstructs a dense Gaussian cloud with rich per-point features, and compresses it into a structured voxel-latent representation using a local voxel transformer that retains only the top-opacity Gaussians per voxel as cross-attention queries. At edit time, the user's images are tokenized and used to update the fast weights of TTT layers inside the refinement network; those adapted weights then transform the

What carries the argument

The load-bearing mechanism is the test-time training (TTT) layer, positioned inside each block of a transformer-based refinement network. Each TTT layer learns a small mapping from keys to values via gradient descent on a self-supervised loss over the tokens of the newly supplied 2D edit images; once the 'fast weights' are updated, the same mapping is applied to both image tokens and the voxel latents that represent the 3D asset. This lets the network absorb the content of an edit into a few adaptable parameters rather than regenerating the asset. Carrying the representation is the compact voxel-latent representation built in a one-time preprocessing step: a voxel transformer compresses a de

Load-bearing premise

The whole pipeline assumes the one-time preprocessing step—rendering the asset from canonical views and compressing the reconstructed Gaussian cloud into voxel latents—faithfully preserves the asset's identity and high-frequency details; if the initial reconstruction misses details, no later refinement can recover them.

What would settle it

Take a set of 3D assets outside the training distribution (e.g., highly intricate mechanical parts or dense foliage), run the preprocessing, and compare a novel-view rendering of the original splat against a rendering of the Stage-I compact latents. If the PSNR between the two drops sharply on such assets while the TTT refinement shows no ability to recover the lost detail, the identity-preserving claim fails. Concretely, measure the novel-view PSNR of the compact representation against the original asset before any edit is applied.

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

If this is right

  • Artists can refine a generated 3D asset by zooming in, painting, or recoloring in a 2D viewport, and see the result in about 0.3 seconds per local edit.
  • Because only fast weights change during an edit, the original asset's geometry and fine structure remain intact; the editing loop is continuous and iterative rather than a one-shot regeneration.
  • A single trained architecture performs multiple appearance-editing tasks—local detail refinement, paint-over, and global relighting—without task-specific fine-tuning.
  • The reported timings (about 0.06 s for global updates, 0.3 s for local edits, plus a one-time ~1.5 s preprocessing per asset) make the approach practical for real-time creative workflows.

Where Pith is reading between the lines

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

  • A likely next step the authors do not take is geometry editing: the method updates appearance attributes only, so extending the voxelized latent to also carry displacement or deformation parameters could unlock structural edits in the same TTT framework.
  • The dependence on a large reconstruction model for the initial feature-rich Gaussian cloud suggests the approach inherits that model's training distribution; assets with very unusual topology or extreme close-up detail may need a higher-resolution or multi-scale voxelization to avoid losing information in the Stage-I compression.
  • One testable extension is to treat the number of TTT gradient steps at inference as a control knob: fewer steps for subtle edits, more steps for large appearance changes, potentially giving users an 'edit strength' slider.
  • Since the editing signal is applied via fast weights shared across the scene, the same architecture could likely be trained to propagate not only relighting but also material or seasonal appearance changes, as long as paired training views can be rendered.

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 / 4 minor

Summary. This paper introduces SplatPainter, a two-stage framework for interactive appearance editing of 3D Gaussian Splatting (3DGS) assets. In Stage I, an input 3DGS asset is rendered from canonical views, reconstructed by GS-LRM into a dense feature-rich Gaussian cloud, and then compressed into a compact voxel latent using a local transformer with top-K opacity-weighted queries. In Stage II, a stack of test-time-training (TTT) layers updates these latents using user-provided 2D edits (local zoom-in refinement, paint-over, or relit views), and a frozen decoder/renderer maps them back to Gaussians. The paper evaluates local refinement and global relighting against optimization, feedforward, and generative baselines, reporting interactive speeds (0.3 s local, 0.06 s global per edit) and claiming state-of-the-art quality.

Significance. The proposed workflow is timely and practically motivated: existing optimization-based editing is slow, and diffusion-based editing is destructive. The TTT-based state-aware design is a sensible way to inject user edits without per-asset optimization, and the speed advantage is clearly demonstrated in Tables 1 and 2. The paper includes a clear architecture diagram, an ablation of the compression design, and a generalization test on Trellis assets. If the quality claims were supported, this would be a notable contribution. However, as reported, the central claim of superior quality in local refinement is not supported by the paper's own numbers, which is a load-bearing inconsistency. The global relighting results are stronger and support the usefulness of the approach.

major comments (3)
  1. [Section 4.2, Table 1] The text states 'our method achieves the best visual and quantitative performance among all baselines.' Table 1 shows Ours (stage II) has PSNR 21.07 vs SRGS 21.01, but SSIM 0.5780 vs 0.6759 (SRGS) and 0.6553 (3DGS), and LPIPS 0.3673 vs 0.3540 (SRGS) and 0.3399 (3DGS). On two of three metrics the optimization baselines are substantially better, and the PSNR gap is only 0.06 dB. This contradicts the abstract's and Section 1's claim of 'significantly outperforms existing methods in both quality and speed.' The authors should either correct the claim, add statistical significance tests/error bars, or provide a user study/perceptual evaluation before the claim can stand.
  2. [Section 3.1, Table 3, Fig. 2] The identity-preservation claim relies on the GS-LRM reconstruction and the top-K opacity compression retaining all details. Table 3 quantifies compression loss only on TexVerse/Objaverse data. For assets outside this distribution (real captures, non-object-centric scenes), the LRM may fail to produce a faithful dense cloud, and Stage II cannot recover details lost in Stage I. Please add an out-of-distribution evaluation or explicitly qualify the claim to in-distribution assets.
  3. [Section 4.1, Tables 1-2] No error bars, confidence intervals, or per-example variance are reported. Given that the PSNR difference between Ours and SRGS in Table 1 is 0.06 dB, and the qualitative statements are 'on-par or better,' statistical support is essential for the 'significantly outperforms' claim. Similarly, Table 2 should report variance across the 370 evaluation examples.
minor comments (4)
  1. [Section 4.2 and Supplementary A.2] The text says 'stage I takes about 0.6 second,' but Table 1 reports 1.54 s and the supplementary says ~1.5 s. Please harmonize these numbers.
  2. [Section 3.2, Fig. 3] The description does not specify whether the fast weights W are retained across sequential edits or re-initialized for each new edit. Since 'continuous editing' is a core claim, clarify the state-persistence semantics.
  3. [Eq. (4), Section 3.2] The adaptation loss is written as an expectation; the concrete mini-batching and number of gradient steps appear only in the implementation details (5 Muon steps). Consider moving that detail into the main text for reproducibility.
  4. [Section 4.4 and Abstract] The abstract emphasizes a 'single architecture' for diverse tasks, but Section A.1 states that two separate Stage II models are trained (local and global). The architecture is unified, but the trained models are task-specific; please state this clearly.

Circularity Check

0 steps flagged

No significant circularity: the editing pipeline is trained end-to-end against ground-truth edited renderings, and the self-citations are reusable components/baselines rather than load-bearing premises.

full rationale

The derivation chain is not circular under the stated criteria. Stage I compresses the GS-LRM output into voxel latents, and its identity-preservation is empirically checked in Table 3 against the uncompressed Pre-Compression baseline (34.02 vs 34.19 PSNR), rather than assumed from the cited GS-LRM paper. Stage II is trained with the image reconstruction loss of Eq. (8), comparing rendered refined Gaussians to ground-truth edited images on held-out TexVerse data; the predicted Gaussian updates are therefore not equivalent to the training inputs by construction. The self-citations (GS-LRM [94], TTT [97], ReLitLRM [96]) are used as network backbones, training techniques, and comparison baselines, not as uniqueness theorems or unverified justifications of the central claim; they are independently published, code-reproduced systems. The apparent mismatch between the Section 4.2 statement that the method 'achieves the best visual and quantitative performance' and Table 1's SSIM/LPIPS numbers for local refinement is a correctness/reporting inconsistency, not a circularity, and is not scored here. No step reduces, by the paper's own equations or by self-citation, to its own inputs.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The paper does not introduce new physical entities or conserved quantities. It depends critically on several empirical assumptions: GS-LRM's feature generality, the fidelity of LRM reconstruction, the compression's detail retention, and the cross-domain transfer of TTT fast weights. These are all plausible but none are proven by the ablation suite alone, especially the assumption that LRM reconstruction preserves the fine details that local refinement aims to add.

free parameters (5)
  • top-K compression ratio = K = max(⌊0.25 N_k⌋, 1)
    The 25% opacity-weighted query per voxel is introduced as an empirical choice ('In practice, we find...') with no sensitivity analysis; the central identity-preservation claim rests on this compression ratio.
  • perceptual loss weight λ_perc = 0.5
    Set for all experiments; no ablation or justification given.
  • number of TTT Muon gradient steps = 5
    Chosen following [97]; no ablation is shown for edit quality vs number of steps.
  • voxel grid resolution = 128^3
    Chosen for memory/compute balance; no ablation on resolution.
  • number of TTT layers in Stage II = 8
    Architecture hyperparameter; no ablation.
axioms (5)
  • domain assumption GS-LRM features trained on Objaverse/TexVerse are sufficiently general to represent arbitrary input 3DGS assets and their fine textures.
    The whole pipeline renders the input asset and re-reconstructs it via GS-LRM (Section 3.1). If LRM drops details for out-of-distribution assets, the refinement cannot recover them.
  • domain assumption The LRM-reconstructed dense Gaussian cloud is 'identity-preserving' and a faithful copy of the original asset.
    Stated in Section 3.1 ('a high-fidelity, identity-preserving copy'); used as the base for both local and global edits.
  • domain assumption Top-K opacity-based query selection preserves perceptually significant details after compression.
    Section 3.1 and ablation Table 3; the ablation compares query types but not K or opacity weighting against other saliency measures.
  • domain assumption TTT fast weights updated on image tokens can be applied to voxel latents to propagate appearance edits in 3D.
    Section 3.2, Eqs. (5)-(7). This cross-domain transfer is the core mechanism and is only validated empirically on the filtered TexVerse set.
  • domain assumption Training on synthetic renders (Objaverse/TexVerse with random HDRI and graffiti overlays) transfers to real user assets.
    Training data described in Section 4.1; generalization is only demonstrated qualitatively on Trellis-generated assets (Fig. 8).

pith-pipeline@v1.3.0-alltime-deepseek · 17235 in / 9222 out tokens · 97732 ms · 2026-08-03T18:21:57.117367+00:00 · methodology

0 comments
read the original abstract

The rise of 3D Gaussian Splatting has revolutionized photorealistic 3D asset creation, yet a critical gap remains for their interactive refinement and editing. Existing approaches based on diffusion or optimization are ill-suited for this task, as they are often prohibitively slow, destructive to the original asset's identity, or lack the precision for fine-grained control. To address this, we introduce SplatPainter, a state-aware feedforward model that enables continuous editing of 3D Gaussian assets from user-provided 2D view(s). Our method directly predicts updates to the attributes of a compact, feature-rich Gaussian representation and leverages Test-Time Training to create a state-aware, iterative workflow. The versatility of our approach allows a single architecture to perform diverse tasks, including high-fidelity local detail refinement, local paint-over, and consistent global recoloring, all at interactive speeds, paving the way for fluid and intuitive 3D content authoring.

Figures

Figures reproduced from arXiv: 2512.05354 by Gordon Wetzstein, Hao Tan, Kai Zhang, Leonidas Guibas, Peng Wang, Wang Yifan, Yang Zheng.

Figure 1
Figure 1. Figure 1: We present SplatPainter, a feedforward method to support interactive and continuous authorship of 3D Gaussian assets through [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview. Given an input 3DGS asset, the framework first performs a one-time preprocessing step. The asset is rendered from multiple views to generate feature-rich inputs from a Gaussian LRM. Stage I compresses this into a compact latent representation via a local transformer. The interactive editing loop in Stage II then iteratively refines this latent representation using new 2D user edits (New input(s))… view at source ↗
Figure 3
Figure 3. Figure 3: TTT operations. Fast weights W are iteratively updated using new input views and subsequently applied to the voxel GS latents after seeing all the new inputs. The residual connections are omitted for clarity. Ximg = {xi}, are projected into key and value vectors us￾ing frozen (slow) projection heads, Wk and Wv: ki = Wkxi , vi = Wvxi , for xi ∈ Ximg. (5) The fast weights W are updated by minimizing Eq. (4))… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative evaluation for local refinement. Given a zoomed-in input view (left), we compare the refined results between baseline methods and ours from a novel view close to zoomed-in view. Our method recovers the details provided by the input zoomed-in views, and produces much sharper features compared to direct reconstruction (GS-LRM [94] or upsampling method (GenDen [50]) and is on-par or better than op… view at source ↗
Figure 5
Figure 5. Figure 5: Global relighting comparison. ReLitLRM takes an environment map as input to synthesize new lighting, while our method leverages a few relit views as direct input. Though not strictly relighting, our method offers a practical path for appearance transfer, achieving consistent shadows, accurate color propagation, and closer alignment with ground truth. shading changes to the entire 3D asset. Baselines. While… view at source ↗
Figure 7
Figure 7. Figure 7: Iterative updating on a global relighting task. From top to bottom: original GS, refined GS with 1-4 input views. Our model picks up partial shading hint (see the first input view) and propagates it coherently to the unseen part (see view 1 in row 2). Generated 3DGS Input Edits Generated 3DGS Input Refined GS Refined GS Original GS Refined GS [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Results of our refinement model on generated 3DGS [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 5 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Information-Regularized Constrained Inversion for Stable Avatar Editing from Sparse Supervision

    cs.CV 2026-04 unverdicted novelty 7.0

    A conditioning-guided constrained inversion method restricts avatar edits to a low-dimensional part-specific subspace and uses an information matrix spectrum from pipeline linearization to predict and ensure stability...

  2. VecSet-Edit: Unleashing Pre-trained LRM for Mesh Editing from Single Image

    cs.CV 2026-02 unverdicted novelty 7.0

    VecSet-Edit is the first method to perform high-fidelity mesh editing from a single image by analyzing and manipulating spatial token subsets in a pre-trained VecSet LRM.

  3. UMI3D: Robust 3D Generation on Unconstrained Multi-Image Inputs via Simultaneous Focus Cross-Attention Routing

    cs.CV 2026-07 conditional novelty 6.0

    Routing each 3D voxel to its most informative conditioning image via a model-intrinsic Voxel Reference Score unlocks robust unconstrained multi-image 3D generation without retraining.

  4. SceneExpander: Text-Guided 3D Scene Expansion via Free-Form View Insertion

    cs.CV 2026-03 conditional novelty 6.0

    A test-time adaptation method integrates a 3D-misaligned, AI-generated inserted view into a reconstructed 3D scene, preserving the captured region while extending it.

  5. VecSet-Edit: Unleashing Pre-trained LRM for Mesh Editing from Single Image

    cs.CV 2026-02 conditional novelty 5.0

    VecSet-Edit edits a mesh by masking and re-denoising a subset of its VecSet latent tokens, preserving untouched regions better than prior voxel-based editors.

Reference graph

Works this paper leans on

104 extracted references · 30 linked inside Pith · cited by 4 Pith papers

  1. [1]

    Building rome in a day.Communications of the ACM, 54(10):105–112, 2011

    Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building rome in a day.Communications of the ACM, 54(10):105–112, 2011. 3

    2011

  2. [2]

    Ed- itp23: 3d editing via propagation of image prompts to multi-view.arXiv preprint arXiv:2506.20652, 2025

    Roi Bar-On, Dana Cohen-Bar, and Daniel Cohen-Or. Ed- itp23: 3d editing via propagation of image prompts to multi-view.arXiv preprint arXiv:2506.20652, 2025. 2

    Pith/arXiv arXiv 2025

  3. [3]

    In- stant3dit: Multiview inpainting for fast editing of 3d ob- jects

    Amir Barda, Matheus Gadelha, Vladimir G Kim, Noam Aigerman, Amit H Bermano, and Thibault Groueix. In- stant3dit: Multiview inpainting for fast editing of 3d ob- jects. InCVPR, pages 16273–16282, 2025. 2

    2025

  4. [4]

    Ti- tans: Learning to memorize at test time.arXiv preprint arXiv:2501.00663, 2024

    Ali Behrouz, Peilin Zhong, and Vahab Mirrokni. Ti- tans: Learning to memorize at test time.arXiv preprint arXiv:2501.00663, 2024. 4

    Pith/arXiv arXiv 2024

  5. [5]

    It’s all connected: A journey through test- time memorization, attentional bias, retention, and online optimization.arXiv preprint arXiv:2504.13173, 2025

    Ali Behrouz, Meisam Razaviyayn, Peilin Zhong, and Va- hab Mirrokni. It’s all connected: A journey through test- time memorization, attentional bias, retention, and online optimization.arXiv preprint arXiv:2504.13173, 2025. 4

    Pith/arXiv arXiv 2025

  6. [6]

    3d-adapter: Geometry- consistent multi-view diffusion for high-quality 3d gener- ation.arXiv preprint arXiv:2410.18974, 2024

    Hansheng Chen, Bokui Shen, Yulin Liu, Ruoxi Shi, Linqi Zhou, Connor Z Lin, Jiayuan Gu, Hao Su, Gordon Wet- zstein, and Leonidas Guibas. 3d-adapter: Geometry- consistent multi-view diffusion for high-quality 3d gener- ation.arXiv preprint arXiv:2410.18974, 2024. 2

    Pith/arXiv arXiv 2024

  7. [7]

    Dge: Direct gaussian 3d editing by consistent multi-view editing.arXiv preprint arXiv:2404.18929, 2024

    Minghao Chen, Iro Laina, and Andrea Vedaldi. Dge: Direct gaussian 3d editing by consistent multi-view editing.arXiv preprint arXiv:2404.18929, 2024. 2, 7

    Pith/arXiv arXiv 2024

  8. [8]

    Shap-editor: Instruction-guided latent 3d editing in sec- onds

    Minghao Chen, Junyu Xie, Iro Laina, and Andrea Vedaldi. Shap-editor: Instruction-guided latent 3d editing in sec- onds. InProceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 26456–26466,

  9. [9]

    Ttt3r: 3d reconstruction as test-time training

    Xingyu Chen, Yue Chen, Yuliang Xiu, Andreas Geiger, and Anpei Chen. Ttt3r: 3d reconstruction as test-time training. arXiv preprint arXiv:2509.26645, 2025. 3

    Pith/arXiv arXiv 2025

  10. [10]

    Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition.Advances in Neural Information Processing Systems, 35:30923–30936, 2022

    Yongwei Chen, Rui Chen, Jiabao Lei, Yabin Zhang, and Kui Jia. Tango: Text-driven photorealistic and robust 3d stylization via lighting decomposition.Advances in Neural Information Processing Systems, 35:30923–30936, 2022. 2

    2022

  11. [11]

    Gaussianeditor: Swift and con- trollable 3d editing with gaussian splatting

    Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xi- aofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huap- ing Liu, and Guosheng Lin. Gaussianeditor: Swift and con- trollable 3d editing with gaussian splatting. InProceedings of the IEEE/CVF conference on computer vision and pat- tern recognition, pages 21476–21485, 2024. 2

    2024

  12. [12]

    Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images

    Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. InEuropean Conference on Computer Vision, pages 370–386. Springer, 2024. 3

    2024

  13. [13]

    Splatformer: Point trans- former for robust 3d gaussian splatting, 2025

    Yutong Chen, Marko Mihajlovic, Xiyi Chen, Yiming Wang, Sergey Prokudin, and Siyu Tang. Splatformer: Point trans- former for robust 3d gaussian splatting, 2025. 2

    2025

  14. [14]

    Text-guided controllable mesh refinement for interactive 3d modeling

    Yun-Chun Chen, Selena Ling, Zhiqin Chen, Vladimir G Kim, Matheus Gadelha, and Alec Jacobson. Text-guided controllable mesh refinement for interactive 3d modeling. InSIGGRAPH Asia 2024 Conference Papers, pages 1–11,

    2024

  15. [15]

    Disco3d: Distilling multi-view consistency for 3d scene editing.arXiv preprint arXiv:2508.01684, 2025

    Yufeng Chi, Huimin Ma, Kafeng Wang, and Jianmin Li. Disco3d: Distilling multi-view consistency for 3d scene editing.arXiv preprint arXiv:2508.01684, 2025. 2

    Pith/arXiv arXiv 2025

  16. [16]

    3d-r2n2: A unified approach for single and multi-view 3d object reconstruction

    Christopher B Choy, Danfei Xu, JunYoung Gwak, Kevin Chen, and Silvio Savarese. 3d-r2n2: A unified approach for single and multi-view 3d object reconstruction. In European conference on computer vision, pages 628–644. Springer, 2016. 3

    2016

  17. [17]

    FlashAttention-2: Faster attention with better par- allelism and work partitioning

    Tri Dao. FlashAttention-2: Faster attention with better par- allelism and work partitioning. InInternational Conference on Learning Representations (ICLR), 2024. 5, 13

    2024

  18. [18]

    3d paintbrush: Local stylization of 3d shapes with cascaded score distillation

    Dale Decatur, Itai Lang, Kfir Aberman, and Rana Hanocka. 3d paintbrush: Local stylization of 3d shapes with cascaded score distillation. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 4473–4483, 2024. 2

    2024

  19. [19]

    Objaverse: A universe of annotated 3d objects

    Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A universe of annotated 3d objects. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13142–13153, 2023. 5, 13

    2023

  20. [20]

    Geometry in style: 3d stylization via sur- face normal deformation

    Nam Anh Dinh, Itai Lang, Hyunwoo Kim, Oded Stein, and Rana Hanocka. Geometry in style: 3d stylization via sur- face normal deformation. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 28456– 28467, 2025. 2

    2025

  21. [21]

    Interactive3d: Create what you want by interactive 3d generation

    Shaocong Dong, Lihe Ding, Zhanpeng Huang, Zibin Wang, Tianfan Xue, and Dan Xu. Interactive3d: Create what you want by interactive 3d generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4999–5008, 2024. 2

    2024

  22. [22]

    Lsd- slam: Large-scale direct monocular slam

    Jakob Engel, Thomas Sch ¨ops, and Daniel Cremers. Lsd- slam: Large-scale direct monocular slam. InEuropean con- ference on computer vision, pages 834–849. Springer, 2014. 3

    2014

  23. [23]

    Preditor3d: Fast and precise 3d shape edit- ing

    Ziya Erkoc ¸, Can G ¨umeli, Chaoyang Wang, Matthias Nießner, Angela Dai, Peter Wonka, Hsin-Ying Lee, and Peiye Zhuang. Preditor3d: Fast and precise 3d shape edit- ing. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 640–649, 2025. 2

    2025

  24. [24]

    3d gaussian splatting as new era: A survey.IEEE Transactions on Visualization and Computer Graphics, 2024

    Ben Fei, Jingyi Xu, Rui Zhang, Qingyuan Zhou, Weidong Yang, and Ying He. 3d gaussian splatting as new era: A survey.IEEE Transactions on Visualization and Computer Graphics, 2024. 2

    2024

  25. [25]

    Srgs: Super-resolution 3d gaussian splatting.arXiv preprint arXiv:2404.10318, 2024

    Xiang Feng, Yongbo He, Yubo Wang, Yan Yang, Wen Li, Yifei Chen, Zhenzhong Kuang, Jianping Fan, Yu Jun, et al. Srgs: Super-resolution 3d gaussian splatting.arXiv preprint arXiv:2404.10318, 2024. 2, 5, 6

    arXiv 2024

  26. [26]

    Svo: Semidirect vi- sual odometry for monocular and multicamera systems

    Christian Forster, Zichao Zhang, Michael Gassner, Manuel Werlberger, and Davide Scaramuzza. Svo: Semidirect vi- sual odometry for monocular and multicamera systems. IEEE Transactions on Robotics, 33(2):249–265, 2016. 3

    2016

  27. [27]

    Kaolin: A pytorch library for accelerating 3d deep learning research

    Clement Fuji Tsang, Maria Shugrina, Jean Francois Lafleche, Or Perel, Charles Loop, Towaki Takikawa, Vis- may Modi, Alexander Zook, Jiehan Wang, Wenzheng Chen, Tianchang Shen, Jun Gao, Krishna Murthy Jataval- labhula, Edward Smith, Artem Rozantsev, Sanja Fidler, Gavriel State, Jason Gorski, Tommy Xiang, Jianing Li, Michael Li, and Rev Lebaredian. Kaolin: ...

  28. [28]

    Textdeformer: Geometry ma- nipulation using text guidance

    William Gao, Noam Aigerman, Thibault Groueix, V ova Kim, and Rana Hanocka. Textdeformer: Geometry ma- nipulation using text guidance. InACM SIGGRAPH 2023 conference proceedings, pages 1–11, 2023. 2

    2023

  29. [29]

    Long short-term memory.Supervised se- quence labelling with recurrent neural networks, pages 37– 45, 2012

    Alex Graves. Long short-term memory.Supervised se- quence labelling with recurrent neural networks, pages 37– 45, 2012. 3

    2012

  30. [30]

    Instruct-nerf2nerf: Edit- ing 3d scenes with instructions

    Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Edit- ing 3d scenes with instructions. InProceedings of the IEEE/CVF international conference on computer vision, pages 19740–19750, 2023. 2, 7

    2023

  31. [31]

    Avatarclip: Zero-shot text-driven generation and animation of 3d avatars.arXiv preprint arXiv:2205.08535, 2022

    Fangzhou Hong, Mingyuan Zhang, Liang Pan, Zhongang Cai, Lei Yang, and Ziwei Liu. Avatarclip: Zero-shot text-driven generation and animation of 3d avatars.arXiv preprint arXiv:2205.08535, 2022. 2

    Pith/arXiv arXiv 2022

  32. [32]

    Lrm: Large reconstruction model for single image to 3d.arXiv preprint arXiv:2311.04400, 2023

    Yicong Hong, Kai Zhang, Jiuxiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d.arXiv preprint arXiv:2311.04400, 2023. 2, 5

    Pith/arXiv arXiv 2023

  33. [33]

    Gsdeformer: Direct, real-time and extensible cage- based deformation for 3d gaussian splatting, 2024

    Jiajun Huang, Shuolin Xu, Hongchuan Yu, and Tong-Yee Lee. Gsdeformer: Direct, real-time and extensible cage- based deformation for 3d gaussian splatting, 2024. 2

    2024

  34. [34]

    Kiri engine: 3d scanner app for iphone, android, and web.https : / / www

    KIRI Innovations. Kiri engine: 3d scanner app for iphone, android, and web.https : / / www . kiriengine . com/, 2024. 2

    2024

  35. [35]

    Perceiver: General perception with iterative attention

    Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. InInternational confer- ence on machine learning, pages 4651–4664. PMLR, 2021. 3

    2021

  36. [36]

    Instruct 3d-to-3d: Text instruction guided 3d-to-3d conversion.arXiv preprint arXiv:2303.15780, 2023

    Hiromichi Kamata, Yuiko Sakuma, Akio Hayakawa, Masato Ishii, and Takuya Narihira. Instruct 3d-to-3d: Text instruction guided 3d-to-3d conversion.arXiv preprint arXiv:2303.15780, 2023. 2

    Pith/arXiv arXiv 2023

  37. [37]

    Learn- ing a multi-view stereo machine.Advances in neural infor- mation processing systems, 30, 2017

    Abhishek Kar, Christian H ¨ane, and Jitendra Malik. Learn- ing a multi-view stereo machine.Advances in neural infor- mation processing systems, 30, 2017. 3

    2017

  38. [38]

    Lattice: Learn- ing to efficiently compress the memory.arXiv preprint arXiv:2504.05646, 2025

    Mahdi Karami and Vahab Mirrokni. Lattice: Learn- ing to efficiently compress the memory.arXiv preprint arXiv:2504.05646, 2025. 4

    arXiv 2025

  39. [39]

    3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42(4), 2023

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42(4), 2023. 1, 2, 5, 6, 13

    2023

  40. [40]

    Decomposing nerf for editing via feature field distil- lation.Advances in neural information processing systems, 35:23311–23330, 2022

    Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitz- mann. Decomposing nerf for editing via feature field distil- lation.Advances in neural information processing systems, 35:23311–23330, 2022. 2

    2022

  41. [41]

    Disr-nerf: Diffusion-guided view-consistent super-resolution nerf

    Jie Long Lee, Chen Li, and Gim Hee Lee. Disr-nerf: Diffusion-guided view-consistent super-resolution nerf. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 20561–20570, 2024. 2

    2024

  42. [42]

    Cmd: Controllable multiview diffu- sion for 3d editing and progressive generation

    Peng Li, Suizhi Ma, Jialiang Chen, Yuan Liu, Congyi Zhang, Wei Xue, Wenhan Luo, Alla Sheffer, Wenping Wang, and Yike Guo. Cmd: Controllable multiview diffu- sion for 3d editing and progressive generation. InProceed- ings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers, pages 1–10, 2025. 2, 7

    2025

  43. [43]

    Advances in 3d generation: A survey.arXiv preprint arXiv:2401.17807, 2024

    Xiaoyu Li, Qi Zhang, Di Kang, Weihao Cheng, Yiming Gao, Jingbo Zhang, Zhihao Liang, Jing Liao, Yan-Pei Cao, and Ying Shan. Advances in 3d generation: A survey.arXiv preprint arXiv:2401.17807, 2024. 1

    Pith/arXiv arXiv 2024

  44. [44]

    Make-your-3d: Fast and consistent subject-driven 3d content generation

    Fangfu Liu, Hanyang Wang, Weiliang Chen, Haowen Sun, and Yueqi Duan. Make-your-3d: Fast and consistent subject-driven 3d content generation. InEuropean Confer- ence on Computer Vision, pages 389–406. Springer, 2024. 2

    2024

  45. [45]

    Inc. Luma AI. Luma ai.https://lumalabs.ai/,

  46. [46]

    Project new depth: Go from 2d to 3d in one click — adobe max sneaks.https: //www.youtube.com/watch?v=1n_p3a- x3yE,

    ´Elie Michel and Adobe. Project new depth: Go from 2d to 3d in one click — adobe max sneaks.https: //www.youtube.com/watch?v=1n_p3a- x3yE,

  47. [47]

    Text2mesh: Text-driven neural stylization for meshes

    Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13492– 13502, 2022. 2

    2022

  48. [48]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis.Communications of the ACM, 65(1):99–106, 2021

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis.Communications of the ACM, 65(1):99–106, 2021. 2

    2021

  49. [49]

    Orb-slam: A versatile and accurate monocular slam system.IEEE transactions on robotics, 31(5):1147–1163,

    Raul Mur-Artal, Jose Maria Martinez Montiel, and Juan D Tardos. Orb-slam: A versatile and accurate monocular slam system.IEEE transactions on robotics, 31(5):1147–1163,

  50. [50]

    Generative densifi- cation: Learning to densify gaussians for high-fidelity gen- eralizable 3d reconstruction

    Seungtae Nam, Xiangyu Sun, Gyeongjin Kang, Younggeun Lee, Seungjun Oh, and Eunbyung Park. Generative densifi- cation: Learning to densify gaussians for high-fidelity gen- eralizable 3d reconstruction. InProceedings of the Com- puter Vision and Pattern Recognition Conference, pages 26683–26693, 2025. 2, 5, 6

    2025

  51. [51]

    Ed-nerf: Efficient text-guided editing of 3d scene with latent space nerf.arXiv preprint arXiv:2310.02712, 2023

    Jangho Park, Gihyun Kwon, and Jong Chul Ye. Ed-nerf: Efficient text-guided editing of 3d scene with latent space nerf.arXiv preprint arXiv:2310.02712, 2023. 2

    Pith/arXiv arXiv 2023

  52. [52]

    Supersplat: 3d gaussian splat editor.https: //supersplat.com/, 2024

    PlayCanvas. Supersplat: 3d gaussian splat editor.https: //supersplat.com/, 2024. 2

    2024

  53. [53]

    Dreamfusion: Text-to-3d using 2d diffusion.arXiv preprint arXiv:2209.14988, 2022

    Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Milden- hall. Dreamfusion: Text-to-3d using 2d diffusion.arXiv preprint arXiv:2209.14988, 2022. 2

    Pith/arXiv arXiv 2022

  54. [54]

    Tailor3d: Customized 3d assets editing and generation with dual-side images.arXiv preprint arXiv:2407.06191, 2024

    Zhangyang Qi, Yunhan Yang, Mengchen Zhang, Long Xing, Xiaoyang Wu, Tong Wu, Dahua Lin, Xihui Liu, Ji- aqi Wang, and Hengshuang Zhao. Tailor3d: Customized 3d assets editing and generation with dual-side images.arXiv preprint arXiv:2407.06191, 2024. 2

    Pith/arXiv arXiv 2024

  55. [55]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021. 2

    2021

  56. [56]

    Elevating 3d models: High- quality texture and geometry refinement from a low-quality model

    Nuri Ryu, Jiyun Won, Jooeun Son, Minsu Gong, Joo-Haeng Lee, and Sunghyun Cho. Elevating 3d models: High- quality texture and geometry refinement from a low-quality model. InProceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers, pages 1–12, 2025. 2

    2025

  57. [57]

    Lin- ear transformers are secretly fast weight programmers

    Imanol Schlag, Kazuki Irie, and J ¨urgen Schmidhuber. Lin- ear transformers are secretly fast weight programmers. In International conference on machine learning, pages 9355–

  58. [58]

    Structure-from-motion revisited

    Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 4104–4113, 2016. 3

    2016

  59. [59]

    Bidirectional recur- rent neural networks.IEEE transactions on Signal Process- ing, 45(11):2673–2681, 1997

    Mike Schuster and Kuldip K Paliwal. Bidirectional recur- rent neural networks.IEEE transactions on Signal Process- ing, 45(11):2673–2681, 1997. 3

    1997

  60. [60]

    V ox-e: Text-guided voxel editing of 3d ob- jects

    Etai Sella, Gal Fiebelman, Peter Hedman, and Hadar Averbuch-Elor. V ox-e: Text-guided voxel editing of 3d ob- jects. InProceedings of the IEEE/CVF international con- ference on computer vision, pages 430–440, 2023. 2

    2023

  61. [61]

    Glu variants improve transformer.arXiv preprint arXiv:2002.05202, 2020

    Noam Shazeer. Glu variants improve transformer.arXiv preprint arXiv:2002.05202, 2020. 4, 13

    Pith/arXiv arXiv 2002

  62. [62]

    Su- pergaussian: Repurposing video models for 3d super reso- lution

    Yuan Shen, Duygu Ceylan, Paul Guerrero, Zexiang Xu, Niloy J Mitra, Shenlong Wang, and Anna Fr ¨uhst¨uck. Su- pergaussian: Repurposing video models for 3d super reso- lution. InEuropean Conference on Computer Vision, pages 215–233. Springer, 2024. 2

    2024

  63. [63]

    Blending-nerf: Text-driven local- ized editing in neural radiance fields

    Hyeonseop Song, Seokhun Choi, Hoseok Do, Chul Lee, and Taehyeong Kim. Blending-nerf: Text-driven local- ized editing in neural radiance fields. InProceedings of the IEEE/CVF international conference on computer vi- sion, pages 14383–14393, 2023. 2

    2023

  64. [64]

    Neuralrecon: Real-time coherent 3d re- construction from monocular video

    Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d re- construction from monocular video. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15598–15607, 2021. 3

    2021

  65. [65]

    Test-time training with self- supervision for generalization under distribution shifts

    Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self- supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229–

  66. [66]

    Learning to (learn at test time): Rnns with expressive hidden states.arXiv preprint arXiv:2407.04620, 2024

    Yu Sun, Xinhao Li, Karan Dalal, Jiarui Xu, Arjun Vikram, Genghan Zhang, Yann Dubois, Xinlei Chen, Xiaolong Wang, Sanmi Koyejo, et al. Learning to (learn at test time): Rnns with expressive hidden states.arXiv preprint arXiv:2407.04620, 2024. 4

    Pith/arXiv arXiv 2024

  67. [67]

    Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras.Advances in neural information processing systems, 34:16558–16569,

    Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras.Advances in neural information processing systems, 34:16558–16569,

  68. [68]

    Cage-gs: High-fidelity cage based 3d gaussian splatting deformation.arXiv preprint arXiv:2504.12800, 2025

    Yifei Tong, Runze Tian, Xiao Han, Dingyao Liu, Fenggen Yu, and Yan Zhang. Cage-gs: High-fidelity cage based 3d gaussian splatting deformation.arXiv preprint arXiv:2504.12800, 2025. 2

    Pith/arXiv arXiv 2025

  69. [69]

    Instruct-gs2gs: Editing 3d gaussian splats with instructions, 2024

    Cyrus Vachha and Ayaan Haque. Instruct-gs2gs: Editing 3d gaussian splats with instructions, 2024. 2, 7

    2024

  70. [70]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017. 3

    2017

  71. [71]

    Sv3d: Novel multi-view synthesis and 3d generation from a single image using la- tent video diffusion

    Vikram V oleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using la- tent video diffusion. InEuropean Conference on Computer Vision, pages 439–457. Springer, 2024. 2

    2024

  72. [72]

    S2gaussian: Sparse-view super-resolution 3d gaussian splatting.arXiv preprint arXiv:2503.04314, 2025

    Yecong Wan, Mingwen Shao, Yuanshuo Cheng, and Wang- meng Zuo. S2gaussian: Sparse-view super-resolution 3d gaussian splatting.arXiv preprint arXiv:2503.04314, 2025. 2

    Pith/arXiv arXiv 2025

  73. [73]

    Clip-nerf: Text-and-image driven ma- nipulation of neural radiance fields

    Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven ma- nipulation of neural radiance fields. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3835–3844, 2022. 2

    2022

  74. [74]

    Nerf-art: Text-driven neu- ral radiance fields stylization.IEEE Transactions on Visu- alization and Computer Graphics, 30(8):4983–4996, 2023

    Can Wang, Ruixiang Jiang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Nerf-art: Text-driven neu- ral radiance fields stylization.IEEE Transactions on Visu- alization and Computer Graphics, 30(8):4983–4996, 2023. 2

    2023

  75. [75]

    Diffusion models for 3d generation: A survey

    Chen Wang, Hao-Yang Peng, Ying-Tian Liu, Jiatao Gu, and Shi-Min Hu. Diffusion models for 3d generation: A survey. Computational Visual Media, 11(1):1–28, 2025. 1

    2025

  76. [76]

    Inpaintnerf360: Text-guided 3d inpainting on unbounded neural radiance fields.arXiv preprint arXiv:2305.15094, 2(4):9, 2023

    Dongqing Wang, Tong Zhang, Alaa Abboud, and Sabine S¨usstrunk. Inpaintnerf360: Text-guided 3d inpainting on unbounded neural radiance fields.arXiv preprint arXiv:2305.15094, 2(4):9, 2023. 2

    Pith/arXiv arXiv 2023

  77. [77]

    3d reconstruction with spatial memory.arXiv preprint arXiv:2408.16061, 2024

    Hengyi Wang and Lourdes Agapito. 3d reconstruction with spatial memory.arXiv preprint arXiv:2408.16061, 2024. 3

    Pith/arXiv arXiv 2024

  78. [78]

    Test- time regression: a unifying framework for designing se- quence models with associative memory.arXiv preprint arXiv:2501.12352, 2025

    Ke Alexander Wang, Jiaxin Shi, and Emily B Fox. Test- time regression: a unifying framework for designing se- quence models with associative memory.arXiv preprint arXiv:2501.12352, 2025. 2, 4

    Pith/arXiv arXiv 2025

  79. [79]

    Continuous 3d perception model with persistent state

    Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 10510–10522, 2025. 3

    2025

  80. [80]

    Intergsedit: Interactive 3d gaussian splatting editing with 3d geometry-consistent attention prior

    Minghao Wen, Shengjie Wu, Kangkan Wang, and Dong Liang. Intergsedit: Interactive 3d gaussian splatting editing with 3d geometry-consistent attention prior. InProceed- ings of the IEEE/CVF International Conference on Com- puter Vision, pages 26136–26145, 2025. 2

    2025

Showing first 80 references.