Pith. sign in

REVIEW 4 major objections 4 minor 37 references

A closed-form per-token control rule, driven by the directional mismatch between source and target velocity fields, localizes training-free 3D edits while preserving untouched geometry.

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

2026-08-02 00:39 UTC pith:EBXTB4BZ

load-bearing objection TanGO is a clean, cheap per-token steering method for 3D editing, but its localization claim rests on an assumption the paper's own evidence does not actually test. the 4 major comments →

arxiv 2607.14927 v1 pith:EBXTB4BZ submitted 2026-07-16 cs.CV

TanGO: Training-Free 3D Editing via Tangent-Space Guidance and Optimization

Siwoo Lim , Sunjae Yoon , Gwanhyeong Koo , Hyeonseo Yun , Chang D. Yoo This is my paper
classification cs.CV
keywords 3D editingtraining-free editingflow matchingVecSet latent representationper-token steeringoptimal controldirectional discrepancytangent-space guidance
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.

TanGO claims that training-free 3D editing fails when the same steering signal is applied to all latent tokens, because preserved and editable regions get perturbed together. Its central idea is to give each token its own steering strength, computed from the directional disagreement between the source and target velocity fields at that token. The paper shows this per-token gain can be derived as the closed-form solution of a one-step optimal control problem, with no masks, no fine-tuning, and negligible extra compute. If correct, it makes precise text-driven edits—add, remove, replace, pose, style—practical on pretrained VecSet-based 3D generators while keeping unedited geometry intact.

Core claim

On the paper's own terms, the discovery is that editing a 3D asset in a VecSet-based flow generator can be cast as an instantaneous optimal-control problem in the tangent space of the generative dynamics. At each ODE step, the method computes source and target velocity fields for every latent token, normalizes them, and defines a demand d_i(t)=1−cosθ_i(t), the directional discrepancy between the two. The closed-form optimal control is u_i(t)=λ_eff(t) g_i(t) Δv_i(t) with g_i(t) proportional to d_i(t), so tokens whose source and target directions disagree strongly are steered hard, while tokens that already agree are left alone. Mean-gain normalization keeps the overall guidance energy stable

What carries the argument

The load-bearing object is the per-token steering gain g_i(t)=d_i(t)/ρ, with demand d_i(t)=1−cosθ_i(t), where θ_i(t) is the angle between the normalized source and target velocity vectors at token i. The demand is motivated as a scaled von Mises-Fisher negative log-likelihood: when the target direction is far from the source direction, demand is high. The update u_i(t)=λ_eff(t) g_i(t) Δv_i(t) comes from a one-step optimal control problem that maximizes progress toward Δv_i while penalizing control energy and restricting updates to the span of Δv_i. The second mechanism is mean-gain normalization, λ_eff(t)=λ·η/(ḡ(t)+ε), which prevents the average steering strength from drifting over time; wit

Load-bearing premise

The method assumes that a token's directional velocity mismatch between source and target is a reliable indicator of whether that token's local geometric region should be edited: high mismatch means edit, low mismatch means preserve.

What would settle it

A causal perturbation experiment would settle it: for a set of edits, set the demand of individual tokens to zero one at a time and record which mesh regions stop changing. If high-demand tokens include regions that should be preserved, or if low-demand tokens include regions that must change, the directional mismatch is not the right localization signal. An oracle-mask comparison—using ground-truth edit masks instead of d_i(t)—would quantify how much of the gain comes from the demand design versus the per-token control structure itself.

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

If this is right

  • Text-driven edits can be applied to existing VecSet-based 3D generators without retraining or external masks: the per-token gain automatically suppresses changes in preserved regions.
  • Rigid edits (add, replace, remove) and non-rigid edits (pose, style) share the same control rule, and the paper reports strong results on both.
  • Because the control is computed from the two velocity evaluations already required, the added computational cost is small compared with prior flow-based editing methods.
  • The same mechanism transfers across multiple VecSet-based generators with different token budgets, suggesting it is not tied to one backbone.
  • Gain normalization is necessary: without it, steering strength drifts over timesteps and leads to under-editing or over-editing.
  • On a mask-aware 3D benchmark, the per-token approach preserves unedited geometry better than baselines, as measured by Chamfer distance and normal consistency inside preserved regions.

Where Pith is reading between the lines

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

  • This suggests the same demand signal could serve as a general mask-free localization prior for other tokenized generative models—for example, locating which tokens in a latent video or audio diffusion model must change for an instruction.
  • Because d_i(t) is derived purely from velocity directions, one could use it as an interpretability tool: plotting per-token demand over steps may reveal which parts of the object the model believes are being edited.
  • A direct comparison against oracle-mask per-token control would separate the contribution of the demand design from the per-token steering structure; the paper does not run that ablation.
  • The observed token locality is measured with an attention-based proxy, not by causal perturbation; if that proxy is imperfect, refining d_i(t) with a learned or mask-based prior could close the gap—a natural extension the paper leaves implicit.

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. The paper proposes TanGO, a training-free method for localized text-driven editing of VecSet-based 3D flow-matching models (e.g., Hunyuan3D 2.1, TripoSG). The method formulates per-token editing as an instantaneous one-step optimal control problem, derives a closed-form control input u_i(t) = λ_eff(t) g_i(t) Δv_i(t), and sets the per-token demand d_i(t) = 1 − cos θ_i(t) from the direction mismatch between source and target velocity fields. A mean-gain normalization is introduced to stabilize global edit strength. The authors introduce a new benchmark, TanGOEdit, and report state-of-the-art results against MVEdit, EditP23, FlowEdit, and AnchorFlow on CLIP-I, CLIP-T, and DINO-I, with additional mask-aware 3D metrics in the supplementary material.

Significance. If the central localization claim is valid, TanGO is a practically useful contribution: it is training-free, mask-free, computationally light (28.73 s per edit), transfers across two VecSet-based backbones, and the closed-form derivation in Section 4 and Appendix B is correct as stated. The release of code and the inclusion of a new benchmark are also assets. However, the paper's headline claim — that per-token directional discrepancy identifies tokens whose geometric region should be edited — rests on an attention-based proxy that the authors themselves state does not directly measure geometric deformation under token perturbation. In addition, the state-of-the-art comparison is evaluated on a self-curated benchmark on which the default hyperparameters were tuned, and no significance testing is reported. These issues are load-bearing for the claimed contribution and need to be addressed before publication.

major comments (4)
  1. [Section 4.2, Eq. (10); Appendix A.1] The core mechanism of the paper is that d_i(t) = 1 − cos θ_i(t) localizes edits: high-demand tokens should lie in the edit region and low-demand tokens in preserved regions. The only locality evidence is the attention-based proxy in Appendix A.1, which explicitly 'does not directly measure geometric deformation under explicit token perturbation' (Remark, Appendix A.1). The mask-aware evaluation in Supp. E.1 reports final geometry metrics on 150 annotated samples but never measures whether the per-token gain map (e.g., accumulated ||u_i(t)|| or d_i(t)) coincides with the annotated edit mask. Without this measurement, the ablation gains in Supp. Table 3 cannot be attributed to the proposed localization mechanism. Please report a direct token-level agreement measure (e.g., mask IoU, precision/recall) on the annotated samples, or a causal token-perturbation experiment that shows high-demand
  2. [Section 5.3, Table 1; Section 5.4, Fig. 8] The headline SOTA claim is evaluated on TanGOEdit, a self-curated benchmark of 100 samples (Supp. C). The default λ and η are selected by sweeping the same TanGOEdit benchmark (Fig. 8(b)). No error bars, confidence intervals, or significance tests are provided for Table 1 or Supp. Table 2. This creates a tuning-on-test-set risk for the central quantitative claim. Please report variability across samples (e.g., bootstrap confidence intervals or per-category breakdowns), use an external benchmark for the main comparison where possible, and show that the chosen hyperparameters are not overfit to the evaluation set.
  3. [Section 4.3, Eqs. (11)–(12)] After mean-gain normalization, λ and η are not separately identifiable. With ε → 0, Eq. (12) gives u_i(t) = λη (d_i(t)/mean_d(t)) Δv_i(t), so only the product λη matters. This makes the two-dimensional (λ, η) sweep in Fig. 8(b) effectively one-dimensional, and it calls into question the interpretation of η as a 'target energy level' that is independent of λ. The same cancellation applies to the vMF concentration κ: since d_i appears only through ratios after normalization, the vMF derivation in Eqs. (8)–(10) does not determine the demand shape beyond a monotone function of 1 − cos θ. Please clarify which parameters are identifiable and what the vMF derivation adds beyond choosing a monotone discrepancy.
  4. [Section 5.1, Implementation Details; Eq. (7)] The source and target classifier-free guidance scales differ substantially (s_src = 3.5, s_tar = 7.5). Since v_tar is produced with much stronger guidance, the cosine mismatch in Eq. (7) may reflect a global shift in the text-conditioned velocity field rather than a spatially localized edit demand. This potential confound is not analyzed. Please report an experiment with matched guidance scales, or show that the d_i(t) map remains localized over a range of s_tar values, to support the spatial localization interpretation.
minor comments (4)
  1. [Eq. (6) vs. Eq. (11)] The symbol ε is used both as a normalization stabilizer in Eq. (6) and as a division-protection term in Eq. (11). Use different symbols (e.g., ε_norm and ε_gain) to avoid confusion.
  2. [Appendix B.5] The appendix correctly notes that the regularized normalization in Eq. (6) is not exactly unit-norm. This is an important implementation detail, but the main text should at least mention that the vMF derivation applies to the idealized unit vectors, while the practical demand is a regularized approximation.
  3. [Fig. 2] Panel labels (c) and (d) appear to refer to CLIP-I and CLIP-T, but the caption and text could be clearer about which panel corresponds to which metric. Please check the labels.
  4. [Appendix F] The limitation on thin structures and the VAE bottleneck is presented only in the supplementary material. Since these are directly relevant to the method's applicability, a brief statement in the main text (Section 5 or the conclusion) would help readers calibrate expectations.

Circularity Check

0 steps flagged

No circular derivation: the control law is a stated closed-form function of source/target velocities, and the vMF step is an explicit identity, not a hidden reuse of the result.

full rationale

TanGO's derivation chain is self-contained. The control law follows from the explicit concave objective (Eq. 4): maximizing d_i<u_i,Δv_i> − (ρ/2)||u_i||^2 over u_i∈span(Δv_i) gives the closed-form u_i=(d_i/ρ)Δv_i (Eq. 5), where d_i is an input weighting, not a fitted output. The paper then chooses d_i=1−cosθ_i by defining the vMF negative log-likelihood (Eqs. 8–10); this is a mathematical identity (−log vMF = const + κ(1−cosθ)) and is explicitly labeled as a definition ('We therefore define the demand...'), so it does not secretly embed the evaluation target. The claimed localization rests on an empirical assumption supported by an attention proxy; Supp. A.1 candidly remarks the metric 'does not directly measure geometric deformation under explicit token perturbation,' which is a validity limitation rather than a circular step. No load-bearing self-citation is used: the authors' prior works [18,34] appear only as related-work context, not as justification for the control law or the vMF choice. The only benchmark concern is that λ and η are selected by sweeping the same TanGOEdit benchmark used for final comparison (Fig. 8b); this is a test-set tuning weakness that can bias the reported SOTA numbers, but it is not a reduction of the derivation to its inputs, and TanGO's update rule remains an explicit function of the stated inputs. Under the specified criteria, no circular step can be exhibited.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The method introduces no new physical or latent entities. Its free parameters are limited to global control strengths, but those strengths are tuned on the same benchmark used for evaluation. The main load-bearing assumptions are about token-level locality and the semantic meaning of velocity-direction mismatch.

free parameters (3)
  • global guidance scale λ = 5.0
    Chosen as the default after a hyperparameter sweep on TanGOEdit; CLIP-I and CLIP-T peak around λ=5 in Fig. 8(b).
  • target mean-gain energy η = 0.2
    Selected together with λ in the same sweep; η=0.2 is reported as the optimal combination.
  • classifier-free guidance scales s_src, s_tar = 3.5, 7.5
    Fixed implementation choices stated in Section 5.1; not swept but influence edit strength.
axioms (5)
  • domain assumption VecSet tokens exhibit strong spatial locality, so each token can be edited independently enough for per-token control.
    Section 3.3 and Appendix A provide an attention-based locality proxy, but the editing method depends on this being causally true under perturbation, which is not directly measured.
  • domain assumption The velocity difference Δv = v_tar − v_src is the correct tangent direction for semantic editing.
    Used in Eq. (4) and throughout the method; inherited from prior training-free editing literature.
  • domain assumption Directional disagreement 1−cosθ between normalized source and target velocities identifies which tokens should be edited.
    Central to the demand design in Section 4.2; if disagreement does not align with edit location, the gain suppresses the wrong tokens.
  • ad hoc to paper A vMF likelihood is an appropriate model for target velocity directions around source directions, and κ/ρ can be absorbed into λ.
    The vMF derivation is motivation only; the final demand is equivalent to cosine distance, so the model adds no falsifiable constraint.
  • domain assumption The pretrained 3D VAE reconstructs edited latents with enough fidelity for the edited mesh to be usable.
    Acknowledged in Appendix F as a bottleneck that bounds detail preservation.

pith-pipeline@v1.3.0-alltime-deepseek · 19993 in / 8471 out tokens · 89097 ms · 2026-08-02T00:39:33.243414+00:00 · methodology

0 comments
read the original abstract

While recent flow-matching 3D generative models (e.g., VecSet) adopt structured representations, their tokens share global context, causing conventional training-free editing to suffer from semantic artifacts such as collapsed preserved regions or incomplete transformations. To address this, we propose TanGO, a training-free framework that enables adaptive per-token steering in the tangent space of generative dynamics. To realize this selective control, we formulate a one-step optimal control rule and determine the strength of each token's control signal using a von Mises-Fisher inspired directional discrepancy derived from the source and target velocity fields. Experiments show that TanGO substantially reduces structural artifacts and achieves state-of-the-art performance, outperforming existing 3D editing baselines. The code is publicly available at https://github.com/siw00-lim/TanGO.

Figures

Figures reproduced from arXiv: 2607.14927 by Chang D. Yoo, Gwanhyeong Koo, Hyeonseo Yun, Siwoo Lim, Sunjae Yoon.

Figure 1
Figure 1. Figure 1: Overview of 3D Editing Results. TanGO achieves precise localized edits across diverse categories, preserving unedited geometry and source identity. Abstract. While recent flow-matching 3D generative models (e.g., Vec￾Set) adopt structured representations, their tokens share global context, causing conventional training-free editing to suffer from semantic arti￾facts such as collapsed preserved regions or i… view at source ↗
Figure 2
Figure 2. Figure 2: (a) Input for text-driven 3D editing. (b) Prior methods often introduce se￾mantic artifacts, such as incomplete edits in target regions or structural degradation in preserved areas. In contrast, TanGO preserves unedited regions while achieving precise, localized edits. (c-d) CLIP-I and CLIP-T scores across five editing categories show that TanGO consistently improves the balance between localized preservat… view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of 3D editing frameworks [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: (a) visualizes the localized influence of individual tokens in the VecSet model, where the colored areas indicate specific regions activated by each respective token. (b) illustrates the number of active tokens out of a total of 4096 triggered during the inference of each data category shown on the x-axis. (c) presents the locality score distribution for datasets across various categories, quantifying how … view at source ↗
Figure 5
Figure 5. Figure 5: Overview of TanGO. Editing stage: At each ODE step, we compute token￾wise source/target velocities and their difference ∆vi(t). To selectively amplify update signals in editable regions while suppressing those in regions to be preserved, we formu￾late an instantaneous one-step optimal control that yields a per-token steering gain and applies ui(t). The demand di(t) is defined by the token-wise directional … view at source ↗
Figure 6
Figure 6. Figure 6: (a) Comparison between AnchorFlow (global scaling) and TanGO (per-token steering). AnchorFlow exhibits under-editing and unpreserved regions, whereas TanGO yields well-edited results. (b) Token-wise contribution of AnchorFlow across timesteps t, shown via a global weight map and per-token strength map. (c) Per-token strength of TanGO, demonstrating its ability to adaptively assign stronger signals to targe… view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparisons on text-driven 3D editing. [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative and Quantitative Analysis of Parameter Selection [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative Analysis of TanGO Components. [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 1
Figure 1. Figure 1: Per-token steering trajectory. (left) Heatmaps of ∥ui(t)∥ and mesh de￾coding show TanGO has a cleaner trajectory. (right) Naive composition perturbs the preserved regions. pattern in which each active token tends to cover a relatively broad region. In contrast, TripoSG activates nearly all tokens, corresponding to a much denser partition of the 3D surface. This indicates that the two models distribute repr… view at source ↗
Figure 2
Figure 2. Figure 2: Additional qualitative results. TanGO achieves more localized edits [PITH_FULL_IMAGE:figures/full_fig_p030_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Additional qualitative results (continued) [PITH_FULL_IMAGE:figures/full_fig_p031_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Additional qualitative results (continued) [PITH_FULL_IMAGE:figures/full_fig_p032_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Failure cases: The mesh breaks or distorts on thin structures like windmill blades or parasols. VAE fails to accurately reconstruct during the initial encoding and final decod￾ing stages will inevitably be lost. Consequently, the upper bound of unedited detail preservation is strictly dictated by the fidelity of the VAE, rather than the precision of our editing algorithm. Second, TanGO exhibits limited edi… 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

37 extracted references · 1 canonical work pages

  1. [1]

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., Zhong, H., Zhu, Y., Yang, M., Li, Z., Wan, J., Wang, P., Ding, W., Fu, Z., Xu, Y., Ye, J., Zhang, X., Xie, T., Cheng, Z., Zhang, H., Yang, Z., Xu, H., Lin, J.: Qwen2.5-vl technical report (2025),https://arxiv.org/abs/2502.13923

    Pith/arXiv arXiv 2025

  2. [2]

    arXiv preprint arXiv:2506.20652 (2025)

    Bar-On, R., Cohen-Bar, D., Cohen-Or, D.: Editp23: 3d editing via propagation of image prompts to multi-view. arXiv preprint arXiv:2506.20652 (2025)

    Pith/arXiv arXiv 2025

  3. [3]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Barda, A., Gadelha, M., Kim, V.G., Aigerman, N., Bermano, A.H., Groueix, T.: Instant3dit: Multiview inpainting for fast editing of 3d objects. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 16273–16282 (2025)

    2025

  4. [4]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Brooks, T., Holynski, A., Efros, A.A.: Instructpix2pix: Learning to follow image editing instructions. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 18392–18402 (2023)

    2023

  5. [5]

    In: Pro- ceedings of the IEEE/CVF international conference on computer vision

    Cao, M., Wang, X., Qi, Z., Shan, Y., Qie, X., Zheng, Y.: Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In: Pro- ceedings of the IEEE/CVF international conference on computer vision. pp. 22560– 22570 (2023)

    2023

  6. [6]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650–9660 (2021)

    2021

  7. [7]

    arXiv preprint arXiv:2403.12032 (2024)

    Chen, H., Shi, R., Liu, Y., Shen, B., Gu, J., Wetzstein, G., Su, H., Guibas, L.: Generic 3d diffusion adapter using controlled multi-view editing. arXiv preprint arXiv:2403.12032 (2024)

    Pith/arXiv arXiv 2024

  8. [8]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Chen, R., Chen, Y., Jiao, N., Jia, K.: Fantasia3d: Disentangling geometry and appearance for high-quality text-to-3d content creation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 22246–22256 (2023)

    2023

  9. [9]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Chen, R., Zhang, J., Liang, Y., Luo, G., Li, W., Liu, J., Li, X., Long, X., Feng, J., Tan, P.: Dora: Sampling and benchmarking for 3d shape variational auto-encoders. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 16251–16261 (2025)

    2025

  10. [10]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Chen, Y., Chen, Z., Zhang, C., Wang, F., Yang, X., Wang, Y., Cai, Z., Yang, L., Liu, H., Lin, G.: Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 21476–21485 (2024)

    2024

  11. [11]

    Advances in Neural Information Processing Systems36, 35799–35813 (2023)

    Deitke, M., Liu, R., Wallingford, M., Ngo, H., Michel, O., Kusupati, A., Fan, A., Laforte, C., Voleti, V., Gadre, S.Y., et al.: Objaverse-xl: A universe of 10m+ 3d objects. Advances in Neural Information Processing Systems36, 35799–35813 (2023)

    2023

  12. [12]

    In: 2022 International Conference on Robotics and Automation (ICRA)

    Downs, L., Francis, A., Koenig, N., Kinman, B., Hickman, R., Reymann, K., McHugh, T.B., Vanhoucke, V.: Google scanned objects: A high-quality dataset of 3d scanned household items. In: 2022 International Conference on Robotics and Automation (ICRA). pp. 2553–2560. Ieee (2022)

    2022

  13. [13]

    In: ACM SIGGRAPH 2023 conference proceedings

    Gao, W., Aigerman, N., Groueix, T., Kim, V., Hanocka, R.: Textdeformer: Ge- ometry manipulation using text guidance. In: ACM SIGGRAPH 2023 conference proceedings. pp. 1–11 (2023)

    2023

  14. [14]

    In: Proceedings of the IEEE/CVF interna- tional conference on computer vision

    Haque,A., Tancik,M.,Efros, A.A., Holynski, A., Kanazawa,A.: Instruct-nerf2nerf: Editing 3d scenes with instructions. In: Proceedings of the IEEE/CVF interna- tional conference on computer vision. pp. 19740–19750 (2023) TanGO 17

    2023

  15. [15]

    In: The Twelfth International Conference on Learning Representations (2023)

    Hong, Y., Zhang, K., Gu, J., Bi, S., Zhou, Y., Liu, D., Liu, F., Sunkavalli, K., Bui, T., Tan, H.: Lrm: Large reconstruction model for single image to 3d. In: The Twelfth International Conference on Learning Representations (2023)

    2023

  16. [16]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Huang, Z., Guo, Y.C., Wang, H., Yi, R., Ma, L., Cao, Y.P., Sheng, L.: Mv- adapter: Multi-view consistent image generation made easy. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16377–16387 (2025)

    2025

  17. [17]

    arXiv preprint arXiv:2506.15442 (2025)

    Hunyuan3D, T., Yang, S., Yang, M., Feng, Y., Huang, X., Zhang, S., He, Z., Luo, D., Liu, H., Zhao, Y., et al.: Hunyuan3d 2.1: From images to high-fidelity 3d assets with production-ready pbr material. arXiv preprint arXiv:2506.15442 (2025)

    Pith/arXiv arXiv 2025

  18. [18]

    In: European Conference on Computer Vision

    Koo, G., Yoon, S., Hong, J.W., Yoo, C.D.: Flexiedit: Frequency-aware latent re- finement for enhanced non-rigid editing. In: European Conference on Computer Vision. pp. 363–379. Springer (2024)

    2024

  19. [19]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Kulikov, V., Kleiner, M., Huberman-Spiegelglas, I., Michaeli, T.: Flowedit: Inversion-free text-based editing using pre-trained flow models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 19721–19730 (2025)

    2025

  20. [20]

    arXiv preprint arXiv:2512.03052 (2025)

    Lai, Z., Zhao, Y., Zhao, Z., Liu, H., Lin, Q., Huang, J., Guo, C., Yue, X.: Lattice: Democratize high-fidelity 3d generation at scale. arXiv preprint arXiv:2512.03052 (2025)

    arXiv 2025

  21. [21]

    arXiv preprint arXiv:2508.19247 (2025)

    Li, L., Huang, Z., Feng, H., Zhuang, G., Chen, R., Guo, C., Sheng, L.: Voxhammer: Training-free precise and coherent 3d editing in native 3d space. arXiv preprint arXiv:2508.19247 (2025)

    Pith/arXiv arXiv 2025

  22. [22]

    IEEE Transactions on Pattern Analysis and Machine Intel- ligence (2025)

    Li, Y., Zou, Z.X., Liu, Z., Wang, D., Liang, Y., Yu, Z., Liu, X., Guo, Y.C., Liang, D., Ouyang, W., et al.: Triposg: High-fidelity 3d shape synthesis using large-scale rectified flow models. IEEE Transactions on Pattern Analysis and Machine Intel- ligence (2025)

    2025

  23. [23]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Liang, Y., Yang, X., Lin, J., Li, H., Xu, X., Chen, Y.: Luciddreamer: Towards high- fidelity text-to-3d generation via interval score matching. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 6517–6526 (2024)

    2024

  24. [24]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Lin, C.H., Gao, J., Tang, L., Takikawa, T., Zeng, X., Huang, X., Kreis, K., Fidler, S., Liu, M.Y., Lin, T.Y.: Magic3d: High-resolution text-to-3d content creation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 300–309 (2023)

    2023

  25. [25]

    Psychology Press, 2 edn

    Macmillan, N.A., Creelman, C.D.: Detection Theory: A User’s Guide. Psychology Press, 2 edn. (2004).https://doi.org/10.4324/9781410611147

    work page doi:10.4324/9781410611147 2004

  26. [26]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Michel, O., Bar-On, R., Liu, R., Ben-Efraim, S., Hanocka, R.: Text2mesh: Text- driven neural stylization for meshes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 13492–13502 (2022)

    2022

  27. [27]

    arXiv preprint arXiv:2209.14988 (2022)

    Poole, B., Jain, A., Barron, J.T., Mildenhall, B.: Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988 (2022)

    Pith/arXiv arXiv 2022

  28. [28]

    In: International conference on machine learning

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748–8763. PMLR (2021)

    2021

  29. [29]

    In: Proceedings of the IEEE/CVF winter conference on applications of computer vision

    Suvorov, R., Logacheva, E., Mashikhin, A., Remizova, A., Ashukha, A., Silvestrov, A., Kong, N., Goka, H., Park, K., Lempitsky, V.: Resolution-robust large mask inpainting with fourier convolutions. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 2149–2159 (2022) 18 S. Lim et al

    2022

  30. [30]

    Advances in neural information processing systems36, 8406–8441 (2023)

    Wang, Z., Lu, C., Wang, Y., Bao, F., Li, C., Su, H., Zhu, J.: Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in neural information processing systems36, 8406–8441 (2023)

    2023

  31. [31]

    Advances in Neural Information Processing Systems37, 121859–121881 (2024)

    Wu, S., Lin, Y., Zhang, F., Zeng, Y., Xu, J., Torr, P., Cao, X., Yao, Y.: Direct3d: Scalable image-to-3d generation via 3d latent diffusion transformer. Advances in Neural Information Processing Systems37, 121859–121881 (2024)

    2024

  32. [32]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Xiang, J., Lv, Z., Xu, S., Deng, Y., Wang, R., Zhang, B., Chen, D., Tong, X., Yang, J.: Structured 3d latents for scalable and versatile 3d generation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 21469–21480 (2025)

    2025

  33. [33]

    arXiv preprint arXiv:2510.15019 (2025)

    Ye, J., Xie, S., Zhao, R., Wang, Z., Yan, H., Zu, W., Ma, L., Zhu, J.: Nano3d: A training-free approach for efficient 3d editing without masks. arXiv preprint arXiv:2510.15019 (2025)

    arXiv 2025

  34. [34]

    In: European Conference on Computer Vision

    Yoon, S., Koo, G., Hong, J.W., Yoo, C.D.: Dni: Dilutional noise initialization for diffusion video editing. In: European Conference on Computer Vision. pp. 180–195. Springer (2024)

    2024

  35. [35]

    ACM Transactions On Graphics (TOG)42(4), 1–16 (2023)

    Zhang, B., Tang, J., Niessner, M., Wonka, P.: 3dshape2vecset: A 3d shape repre- sentation for neural fields and generative diffusion models. ACM Transactions On Graphics (TOG)42(4), 1–16 (2023)

    2023

  36. [36]

    ACM Transactions on Graphics (TOG)43(4), 1–20 (2024)

    Zhang, L., Wang, Z., Zhang, Q., Qiu, Q., Pang, A., Jiang, H., Yang, W., Xu, L., Yu, J.: Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG)43(4), 1–20 (2024)

    2024

  37. [37]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Zhou, Z., Ma, F., Gui, C., Xia, X., Fan, H., Yang, Y., Chua, T.S.: Anchorflow: Training-free 3d editing via latent anchor-aligned flows. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14387– 14397 (2026) Supplementary Material for TanGO: Training-Free 3D Editing via Tangent-space Guidance and Optimization Siwoo ...

    2026