Pith. sign in

REVIEW 3 major objections 5 minor 41 references

GOPI: Generation-Oriented 3D Pose Inference for Furniture Insertion from Single-View RGB-D Indoor Scenes

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Pith's one-line read This paper claims that furniture insertion into single-view RGB-D scenes should be solved pose-first: infer a geometrically plausible 3D placement, then constrain image synthesis with its projection, because 2D-only masks leave physical…

desk verdict A competent pose-first furniture insertion system with a clear evaluation gap: the central claim that 3D conditioning beats 2D-only conditioning is never directly tested. read the letter →

arxiv 2608.06836 v1 pith:6NV75C7Z submitted 2026-08-07 cs.CV

classification cs.CV
keywords furnitureinsertion3DposeinferenceRGB-Dindoorscenesgeometry-guidedimagegenerationdiffusionmodelsiterativerefinementscaleambiguitypose-firstformulation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Inserting new furniture into a single-view RGB-D image is underdetermined when the only constraint is a 2D mask: the mask fixes the image-plane extent of the object but not its physical size or 3D position, so many different placements can explain the same observation. The paper proposes a pose-first solution, GOPI, which first infers a geometrically plausible 3D pose (translation plus yaw) for the furniture from masked scene geometry, an insertion frustum, and multi-view furniture geometry, and then uses the rendered projection of that pose as a pixel-aligned conditioning signal for diffusion-based image generation. On a synthetic indoor benchmark, this two-stage design beats direct point-cloud regression and a non-geometric baseline on geometric feasibility and tolerance-based pose accuracy, and it keeps the synthesized furniture aligned with the projected geometry when furniture scale is varied from 0.7 to 1.6. The paper's central conclusion is that solving 3D placement before appearance synthesis is the right ordering for geometry-consistent furniture insertion.

What carries the argument

The load-bearing object is the pose-first refinement loop with geometric conditioning. Concretely, GOPI combines two parallel point-cloud encoders, one for the environment (scene points plus frustum boundary points, each tagged with signed distances to the four lateral frustum planes) and one for the furniture; environment features attend to furniture features via cross-attention, and two lightweight heads read out a translation-yaw update and a dominant room axis. Refinement works by transforming the environment point cloud with the inverse of each predicted update while keeping furniture in its canonical frame, and the frustum consistency loss penalizes furniture points that fall outside the insertion frustum, with stronger weight for visible violations. The final pose is rendered as a geometry-aligned RGB projection in the scene camera and concatenated into a diffusion model as an additional image-conditioning channel, which anchors generated pixels to the inferred 3D location, scale, and orientation.

What would settle it

Retrain and re-evaluate GOPI using only ground-truth poses that pass all three feasibility criteria (support, collision, frustum), and compare overall feasibility and Proj-Gen IoU against the same baselines; if the pose-first gains shrink or disappear on this filtered set, the central claim is undermined. Alternatively, test on real captured RGB-D rooms with sensor noise, since all results come from synthetic renders and the paper itself lists real-world generalization as an open limitation.

Watch

Extended reading notes

Core claim

The central claim is that for geometry-consistent furniture insertion it is beneficial to first solve 3D placement and then use the inferred geometry to constrain synthesis. GOPI carries out this claim by encoding the observed scene as a masked point cloud plus a frustum scaffold for the user's insertion region, encoding the target furniture as a reconstructed object-centric point cloud, and running cross-attention between the two streams. Pose prediction is formulated as iterative refinement: the furniture stays in a canonical frame while the environment is transformed by the inverse of each predicted pose update, with losses for frustum consistency, support alignment, and room-axis alignment. In the reported experiments, GOPI-Dual reaches 68.0% overall feasibility and 69.5% translation accuracy within 0.3 m, versus 26.3% and 31.8% for direct regression and 43.8% and 58.2% for the vanilla baseline; removing iterative refinement drops overall feasibility to 33.1%. For image generation, the rendered projection of the inferred pose is fed into a diffusion model as a pixel-aligned anchor, and the measured projection-generation IoU stays within a 3.83 percentage-point range over furniture scale factors 0.7, 1.0, 1.3, and 1.6.

Load-bearing premise

The ground-truth furniture placements used as training targets and reference evaluations are not all physically valid: Table I shows about 7% fail the paper's own feasibility criteria, and if those imperfect layouts are systematically biased, the measured feasibility improvements could partly reflect learning the dataset's imperfections rather than genuine geometric understanding.

Editorial extensions

If this is right

  • Pose-first ordering should be adopted in furniture insertion systems, because resolving 3D placement before appearance synthesis removes the scale ambiguity that 2D-only conditioning cannot resolve.
  • Iterative refinement is the main engine of geometric feasibility; without it the overall feasibility rate falls from 68.0% to 33.1%, so one-shot pose regression is not enough under masked single-view observability.
  • An open frustum scaffold derived from a user's coarse mask is a usable spatial constraint even though the geometry inside the mask is unknown, and it measurably improves support, collision, and translation accuracy.
  • Conditioning image synthesis on the rendered projection of the inferred pose keeps the output aligned with the geometry across a scale range, making the generated image a reliable visualization of the proposed 3D placement.
  • Room-axis alignment improves yaw accuracy but creates a step-like yaw error profile, meaning predictions close to dominant room axes are corrected while those far from an axis can be pulled in the wrong direction.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves implicit that the pose-first principle should transfer to other rigid object categories besides furniture, since any rigid object with strong scale constraints suffers the same 2D scale ambiguity; a direct test would be applying GOPI's conditioning to appliances or decorative objects in the same scenes.
  • Because roughly 7% of the ground-truth poses used for training and evaluation are themselves infeasible by the paper's feasibility criteria, the reported gains could partly be learned against imperfect targets; retraining on a feasibility-filtered reference set is a natural check on whether pose-first helps more on genuinely valid layouts.
  • A testable extension is to run the same geometry-guided conditioning without depth input, replacing the scene point cloud with monocular depth estimates; if projection-generation alignment degrades gracefully, the framework could move from RGB-D to RGB-only furniture insertion.
  • The Proj-Gen IoU diagnostic itself could serve as a standard alignment metric for any geometry-conditioned image editing pipeline, independent of GOPI, since it checks whether generated pixels respect an explicit 3D hypothesis.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes GOPI, a two-stage framework for inserting new furniture into indoor scene images from masked single-view RGB-D input. The first stage infers a 3D furniture pose (translation and yaw) by jointly encoding observed scene geometry, an insertion frustum derived from the user mask, and multi-view furniture geometry, with an iterative refinement procedure that updates the environment point cloud rather than the object pose. The second stage renders the inferred pose as a pixel-aligned RGB projection and feeds it as an extra conditioning channel to a diffusion model (IDM-based), alongside the masked scene and generation mask. Experiments are conducted on a new synthetic benchmark Front3D-Insertion built from 3D-FRONT/3D-FUTURE. Pose-side results are reported against a PTv3 regression baseline and a vanilla version of GOPI, with feasibility metrics (support, wall collision, frustum, overall) and pose accuracy metrics (TransAcc, YawAcc), plus ablations. Generation-side results report a Proj-Gen IoU metric across furniture scale factors 0.7, 1.0, 1.3, and 1.6. The paper concludes that solving 3D placement first and then using inferred geometry to constrain synthesis is beneficial for geometry-consistent furniture insertion.

Significance. If the central claim were fully established, the pose-first formulation would be a meaningful contribution: it directly addresses the known scale ambiguity of mask-only 2D conditioning and connects 3D layout reasoning with image generation. The pose-side experiments are reasonably thorough: the GOPI model is compared against a direct-regression baseline and a vanilla iterative baseline, and the ablation in Table II clearly attributes most of the gain to iterative refinement. The paper also introduces a useful evaluation protocol for geometric feasibility and projection–generation alignment, and it is transparent about the filtered subset used for Proj-Gen IoU. However, the load-bearing claim—that 3D conditioning is better than 2D-only conditioning for downstream generation—is not tested by the current experiments. The generation results only measure how well the proposed pipeline follows its own projection conditioning, and no 2D-only control is provided. Thus the pose-side contribution is credible and moderately valuable, but the central motivation and conclusion remain unsupported without the missing control.

major comments (3)
  1. [Section V-F, Fig. 8] The central claim of the paper—summarized in Section VII as 'for geometry-consistent furniture insertion, it is beneficial to first solve 3D placement and then use the inferred geometry to constrain synthesis'—is not directly tested. Section V-F reports Proj-Gen mIoU (Eq. 17) only for the proposed geometry-guided pipeline across scale factors 0.7, 1.0, 1.3, and 1.6. There is no control condition that receives the same masked scene and mask but omits the projected 3D geometry, for example a 2D-only diffusion baseline conditioned on the masked scene plus a binary rectangle or a randomly positioned projection. Without such a control, the observed scale stability could be produced by any pixel-aligned target region, because the reference mask M_proj is derived from the same projection I_proj that is concatenated into the diffusion model as a conditioning channel; the metric then partly measures whether the generator follows the supplied projection, not whether the inferred 3D pose drives the consistency. A concrete additional experiment is needed: compare Proj-Gen mIoU and its sensitivity to scale factors for (a) the proposed geometry-guided conditioning, (b) a 2D-only conditioning using the user mask or its bounding rectangle without the projected 3D geometry, and ideally (c) a projection from a perturbed or random pose. Only then can the conclusion in Section VII be supported.
  2. [Section V-A.2, Table I] The ground-truth layouts used for training and evaluation are themselves partially infeasible: Table I reports GT wall-collision validity of 92.9% and overall feasibility of 92.7%, meaning about 7% of reference poses violate at least one feasibility constraint. The paper acknowledges that GT is 'a reference layout rather than a strict upper bound,' yet the regression and refinement losses are still trained toward these poses and the success metrics are measured against them. This creates a risk that the reported feasibility gains partly reflect learning the dataset's imperfections rather than genuine geometric understanding, especially if the infeasible cases are systematically biased (e.g., objects placed near walls or at particular scales). The authors should analyze the subset of infeasible GT poses separately, report metrics on feasible GT samples only, or train a variant that excludes infeasible references from the loss, to show that the conclusions are robust to this 7% contamination.
  3. [Section V-A.2 and Fig. 8] The scale-stability conclusion is computed on a filtered subset: the Proj-Gen IoU metric is reported on 695 test samples whose GT-scale Proj-Gen IoU exceeds 0.75, as stated in Section V-A.2. Figure 8 and the accompanying statement that 'Proj-Gen mIoU remains relatively stable across the tested furniture scale factors' therefore describe behavior only for already well-aligned samples. Section VII generalizes this to the proposed framework as a whole, which overreaches. The authors should either report the same metric on the full test set (acknowledging the segmentation-reliability trade-off) or explicitly frame the scale-stability result as conditional on the filtered subset, without claiming a general property of the framework.
minor comments (5)
  1. [Table I] The row label 'GOPI-Dual, default97.2 79.2 85.7 68.0 69.562.3' has missing whitespace and formatting; it should read 'GOPI-Dual, default' followed by properly separated numbers.
  2. [Section V-F, Eq. (17)] The definition of Proj-Gen IoU uses Grounded-SAM to obtain the generated furniture mask, but the reliability threshold (GT-scale Proj-Gen IoU > 0.75) is applied to a quantity that is itself a function of the same segmentation pipeline; the dependency of the filtered subset on the metric being evaluated should be stated more explicitly, even though the paper does disclose the filtering.
  3. [Section IV-C(g), Eq. (12)] The schedules α(τ), β(τ), the noise magnitude ε_τ, and the convergence thresholds for terminating refinement are described qualitatively but not specified numerically; since these are free parameters that affect pose accuracy, concrete values or a reference to a standard schedule should be provided for reproducibility.
  4. [Section V-A.1] The dataset construction is based on 3D-FRONT/3D-FUTURE and BlenderProc, which is clearly described, but it would help to state the number of distinct scenes, rooms, and furniture categories included in Front3D-Insertion, as these factors influence the diversity of the benchmark.
  5. [Full text] There are several typos and spacing errors, e.g., 'GOPI-Dual, default97.2' in Table I and 'Proj-Gen mIoU' inconsistently written as 'Proj-Gen IoU' in Eq. (17); a careful proofread is recommended.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor self-reference in the Proj-Gen evaluation metric; the pose-estimation derivation chain is otherwise self-contained.

  1. self definitional [Section IV-D.c and Section V-A.2, Eqs. (16)-(17)]
    "As the geometric reference, we extract the projection mask M_proj from the rendered projection I_proj_o by thresholding valid pixels. The metric is defined as IoU_Proj−Gen = |M_gen ∩ M_proj|/|M_gen ∪ M_proj|. ... The TryonNet branch takes the noisy latent z together with the generation mask m, the masked-scene latent z_masked_s, and the projection latent z_proj_o."

    The reference mask M_proj is derived from I_proj_o (Eq. 16), which is exactly the geometry-aligned projection fed into the diffusion model as conditioning (z_proj_o). The Proj-Gen IoU therefore measures how well the generated furniture mask follows the very projection supplied as a pixel-aligned input, rather than testing whether the inferred 3D pose independently drives consistency. The scale-stability conclusion in Fig. 8 is additionally computed on a filtered 695-sample subset and without a 2D-only generation baseline, so the pose-first benefit is not independently established by this metric. This is a partial self-reference in the evaluation protocol, not a fitted-parameter reduction.

full rationale

The paper's load-bearing pose results are empirical and supervised: GOPI is trained on perturbed reference configurations (Eq. 3) with explicit losses and evaluated on held-out test samples against PTv3 Regression and Vanilla baselines. None of the pose-feasibility or pose-accuracy results reduce by construction to a fitted parameter; the feasibility metrics are computed from predicted geometry, wall maps, and frustum projections, and the training target is the dataset reference pose. The only self-referential element is the Proj-Gen IoU metric, whose reference is the same projection used as a conditioning channel; this makes the generation-alignment result a conditioning-fidelity diagnostic and leaves the central pose-first claim only partially tested on the generation side. The missing 2D-only control and the filtered evaluation subset are experimental-design limitations rather than equation-level circularity. No load-bearing self-citation chain is present: the only co-authored reference (Zero-1-to-3 [10]) is contextual, and no uniqueness theorem or ansatz is imported from the authors' prior work. Overall, the circularity is minor and confined to the evaluation protocol, so a score of 2 is appropriate.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The framework rests on assumptions about input format (known gravity, rigid object geometry, masks derived from real objects), structural priors (Manhattan axes, valid GT layouts), and evaluation choices (filtered subset, hand-set thresholds). The many unspecified hyperparameters (thresholds, margins, schedules) shift explanatory weight onto implementation details not fully disclosed.

free parameters (4)
  • Feasibility thresholds τ_z, τ_col, τ_frus = medium thresholds not specified numerically
    The Overall@M metric depends on hand-set thresholds for support, wall collision, and frustum validity; exact values are not given, so comparisons across settings are hard to interpret.
  • Frustum consistency hinge margin m = not specified
    Appears in Eq. 5 and controls how much furniture may protrude outside the frustum before penalty; chosen by hand and not ablated.
  • Proj-Gen IoU reliability threshold = 0.75
    Post hoc filter selecting 695 test samples with GT-scale Proj-Gen IoU above 0.75; directly shapes the reported image-consistency result.
  • Soft-min temperature τ in room orientation loss = not specified
    Eq. 9 hyperparameter controlling the sharpness of the soft minimum over room axes; value not reported.
assumptions (5)
  • domain assumption An upright scene assumption with known gravity aligned to the vertical axis
    Stated in Section III-B.d; pose is parameterized only by translation and yaw, with no roll or pitch.
  • domain assumption Manhattan-world assumption that most furniture orientations align with dominant room axes
    Invoked in Section IV-B.d for the room orientation head and in Eq. 15 for yaw snapping during refinement.
  • domain assumption The target furniture geometry is rigid and available as a multi-view RGB-D reconstruction
    Section III-B.b; the pipeline requires an object-centric point cloud Po built from separate renderings, which may not be available in real single-image insertion.
  • domain assumption Reference 3D-FRONT layouts are valid enough to serve as training targets and evaluation references
    Table I shows GT itself fails wall collision on 7.1% of samples and overall feasibility on 7.3%; the paper treats GT as a reference rather than an upper bound but still trains toward it.
  • domain assumption Insertion mask construction from GT instance bounding boxes simulates real user intent
    Section V-A.1; masks are generated from the target object's rendered region plus random padding, guaranteeing a real object exists behind every mask, which may not hold for user-provided masks.
invented entities (1)
  • Insertion frustum boundary point set (P_f)
    purpose: A geometric scaffold representing the masked insertion region without needing interior geometry; used in the environment point cloud and the frustum consistency loss.
    Introduced in Section IV-A.b; no external validation that a frustum scaffold is the correct representation of user intent, and performance depends on its construction and margin parameters.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GOPI: Generation-Oriented 3D Pose Inference for Furniture Insertion from Single-View RGB-D Indoor Scenes." pith.science (2026). https://pith.science/paper/6NV75C7Z

@misc{pith2026260806836,
  author       = {Pith},
  title        = {Pith review of: GOPI: Generation-Oriented 3D Pose Inference for Furniture Insertion from Single-View RGB-D Indoor Scenes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6NV75C7Z}},
  note         = {Machine review of arXiv:2608.06836}
}
read the original abstract

We study the problem of inserting new furniture into indoor scene images. Under masked single-view 2D image-plane conditioning, however, the physical scale of the inserted furniture relative to the scene cannot be uniquely determined, making physically grounded furniture placement underdetermined from image evidence alone. We therefore reformulate the task as a combination of 3D pose inference and geometry-guided image generation, where estimating a geometrically plausible 3D placement is essential for reliable synthesis. To this end, we propose a two-stage framework. For 3D placement, we introduce GOPI, a generation-oriented 3D pose inference framework that addresses the underdetermined nature of single-view furniture insertion through data-driven iterative inference, producing geometrically plausible object placements. For image generation, we develop a geometry-guided conditioning strategy that projects the inferred 3D pose into the image plane as a pixel-aligned constraint, enforcing consistency between the synthesized image and the underlying 3D geometry. Experimental results validate the proposed framework from both 3D pose estimation and image synthesis perspectives. For 3D placement, GOPI produces poses with stronger geometric feasibility and better consistency with reference layouts than direct regression and vanilla baselines. For image synthesis, our method preserves alignment with the projected 3D geometry across different furniture scales, showing stable projection-generation alignment across the tested furniture scales.

Figures

Figures reproduced from arXiv: 2608.06836 by the authors.

Figure 1
Figure 1. Ambiguity under 2D vs. scale consistency with 3D guidance. (a) Under 2D-only conditioning, the same masked input can correspond to multiple geometrically plausible insertion results with different object scales, due to inherent scale ambiguity. (b) By introducing 3D reasoning, the ambiguity is reduced, leading to scale-consistent image generation that aligns with the underlying 3D geometry. Furniture differs fundame… view at source ↗
Figure 2
Figure 2. Overview of the furniture insertion task and the GOPI framework. The inputs consist of a masked single-view scene RGB-D observation, obtained by capturing an indoor scene and specifying a user mask, together with a multi-view RGB-D set of the target furniture. Based on these inputs, GOPI constructs scene–frustum geometry and furniture geometry, encodes them with environment and object encoders, performs cross-stream… view at source ↗
Figure 3
Figure 3. Diffusion conditioning with geometry-aligned inputs. The geometry-aware inputs are concatenated along the channel dimension and fed into the diffusion model, where the geometry-aligned projection provides a pixel-aligned spatial constraint that anchors the generated content to the inferred 3D pose. D. Geometry-Guided Image Generation a) Geometry-aligned RGB Projection.: Given the final refined pose p ∗ , we render a… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: further analyzes GOPI under strict, medium, and loose feasibility thresholds. Support validity remains consis￾tently high, whereas wall-collision and frustum validity form the main bottlenecks. Since overall feasibility requires support, collision, and frustum validity…
Figure 5
Figure 5. Figure 5: compares collision and frustum success rates under varying thresholds. GOPI consistently outperforms Vanilla and Overall Collision Frustum Support 0 20 40 60 80 100 Success Rate (%) Strict Medium Loose [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 7
Figure 7. Figure 7: compares GOPI under open-mask and closed-mask settings. Closed masks consistently lead to better performance than open masks. The largest gap appears in TransAcc@0.3m, with a difference of 10.7 pp, followed by Overall@M with a gap of 8.8 pp. This degradation is mainly …
Figure 8
Figure 8. Figure 8: Projection–generation mIoU of geometry-guided image generation under different furniture scale factors. The shaded region shows the range of Proj-Gen mIoU across the tested scale factors, while the dashed line denotes the GT-scale Proj-Gen mIoU computed on the same fil…
Figure 9
Figure 9. Figure 9: Qualitative results under different furniture scale factors. Green overlays denote the projected 2D masks of the inferred 3D furniture geometry. Zoom in for details. 2D masks of the inferred 3D furniture geometry. Across different furniture scales, these projected mask…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 24 canonical work pages

  1. [1]

    High- Resolution Image Synthesis With Latent Diffusion Models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- Resolution Image Synthesis With Latent Diffusion Models,” inCVPR, Jun. 2022, pp. 10 684–10 695

  2. [2]

    SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis,

    D. Podell, Z. English, K. Lacey, A. Blattmann, T. Dockhorn, J. M ¨uller, J. Penna, and R. Rombach, “SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis,” inThe Twelfth International Conference on Learning Representations, 2024

  3. [3]

    RePaint: Inpainting Using Denoising Diffusion Proba- bilistic Models,

    A. Lugmayr, M. Danelljan, A. Romero, F. Yu, R. Timofte, and L. Van Gool, “RePaint: Inpainting Using Denoising Diffusion Proba- bilistic Models,” inCVPR, 2022, pp. 11 461–11 471

  4. [4]

    DiffEdit: Diffusion-based semantic image editing with mask guidance,

    G. Couairon, J. Verbeek, H. Schwenk, and M. Cord, “DiffEdit: Diffusion-based semantic image editing with mask guidance,” inIn- ternational Conference on Learning Representations, 2023

  5. [5]

    SmartBrush: Text and Shape Guided Object Inpainting With Diffusion Model,

    S. Xie, Z. Zhang, Z. Lin, T. Hinz, and K. Zhang, “SmartBrush: Text and Shape Guided Object Inpainting With Diffusion Model,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun. 2023, pp. 22 428–22 437

  6. [6]

    Uni-paint: A Unified Framework for Multimodal Image Inpainting with Pretrained Diffusion Model,

    S. Yang, X. Chen, and J. Liao, “Uni-paint: A Unified Framework for Multimodal Image Inpainting with Pretrained Diffusion Model,” inACM International Conference on Multimedia, 2023, pp. 3190–3199

  7. [7]

    DreamFusion: Text-to-3D using 2D Diffusion,

    B. Poole, A. Jain, J. T. Barron, and B. Mildenhall, “DreamFusion: Text-to-3D using 2D Diffusion,” inInternational Conference on Learning Representations, 2023. [Online]. Available: https://arxiv.org/ abs/2209.14988

  8. [8]

    Magic3D: High-Resolution Text- to-3D Content Creation,

    C.-H. Lin, J. Gao, L. Tang, T. Takikawa, X. Zeng, X. Huang, K. Kreis, S. Fidler, M.-Y . Liu, and T.-Y . Lin, “Magic3D: High-Resolution Text- to-3D Content Creation,” inCVPR, Jun. 2023, pp. 300–309

Show all 41 references
  1. [9]

    Prolific- Dreamer: High-Fidelity and Diverse Text-to-3D Generation with Varia- tional Score Distillation,

    Z. Wang, C. Lu, Y . Wang, F. Bao, C. LI, H. Su, and J. Zhu, “Prolific- Dreamer: High-Fidelity and Diverse Text-to-3D Generation with Varia- tional Score Distillation,” inAdvances in Neural Information Processing Systems, vol. 36, 2023, pp. 8406–8441

  2. [10]

    Zero-1-to-3: Zero-shot One Image to 3D Object,

    R. Liu, R. Wu, B. Van Hoorick, P. Tokmakov, S. Zakharov, and C. V ondrick, “Zero-1-to-3: Zero-shot One Image to 3D Object,” inICCV, Oct. 2023, pp. 9298–9309

  3. [11]

    MVDream: Multi-view Diffusion for 3D Generation,

    Y . Shi, P. Wang, J. Ye, M. Long, K. Li, and X. Yang, “MVDream: Multi-view Diffusion for 3D Generation,” inInternational Conference on Learning Representations, 2024. [Online]. Available: https://arxiv.org/abs/2308.16512

  4. [12]

    Wonder3D: Single Image to 3D using Cross-Domain Diffusion,

    X. Long, Y .-C. Guo, C. Lin, Y . Liu, Z. Dou, L. Liu, Y . Ma, S.-H. Zhang, M. Habermann, C. Theobalt, and W. Wang, “Wonder3D: Single Image to 3D using Cross-Domain Diffusion,” inCVPR, Jun. 2024, pp. 9970– 9980

  5. [13]

    SceneFormer: Indoor Scene Generation with Transformers,

    X. Wang, C. Yeshwanth, and M. Nießner, “SceneFormer: Indoor Scene Generation with Transformers,” in2021 International Conference on 3D Vision (3DV), 2021, pp. 106–115

  6. [14]

    ATISS: Autoregressive Transformers for Indoor Scene Synthesis,

    D. Paschalidou, A. Kar, M. Shugrina, K. Kreis, A. Geiger, and S. Fidler, “ATISS: Autoregressive Transformers for Indoor Scene Synthesis,” in Advances in Neural Information Processing Systems, vol. 34, 2021. [Online]. Available: https://arxiv.org/abs/2110.03675

  7. [15]

    DiffuScene: Denoising Diffusion Models for Generative Indoor Scene Synthesis,

    J. Tang, Y . Nie, L. Markhasin, A. Dai, J. Thies, and M. Nießner, “DiffuScene: Denoising Diffusion Models for Generative Indoor Scene Synthesis,” inCVPR, Jun. 2024, pp. 20 507–20 518

  8. [16]

    LEGO-Net: Learning Regular Rearrangements of Objects in Rooms,

    Q. A. Wei, S. Ding, J. J. Park, R. Sajnani, A. Poulenard, S. Sridhar, and L. Guibas, “LEGO-Net: Learning Regular Rearrangements of Objects in Rooms,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 24 017–24 026. [Online]. Availab...

  9. [17]

    Classifier-Free Diffusion Guidance,

    J. Ho and T. Salimans, “Classifier-Free Diffusion Guidance,” 2022, arXiv: 2207.12598. [Online]. Available: https://arxiv.org/abs/2207.12598

  10. [18]

    LatentPaint: Image Inpainting in Latent Space with Diffusion Models,

    C. Corneanu, R. Gadde, and A. M. Martinez, “LatentPaint: Image Inpainting in Latent Space with Diffusion Models,” in2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). Los Alamitos, CA, USA: IEEE Computer Society, Jan. 2024, pp. 4322–4331. [Online]. Avai...

  11. [19]

    InstructEdit: Improving Automatic Masks for Diffusion-based Image Editing With User Instructions,

    Q. Wang, B. Zhang, M. Birsak, and P. Wonka, “InstructEdit: Improving Automatic Masks for Diffusion-based Image Editing With User Instructions,” 2023, arXiv: 2305.18047. [Online]. Available: https://arxiv.org/abs/2305.18047

  12. [20]

    MAG-Edit: Local- ized Image Editing in Complex Scenarios via Mask-Based Attention- Adjusted Guidance,

    Q. Mao, L. Chen, Y . Gu, Z. Fang, and M. Z. Shou, “MAG-Edit: Local- ized Image Editing in Complex Scenarios via Mask-Based Attention- Adjusted Guidance,” inProceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 6842–6850. 11

  13. [21]

    FlexEdit: Marrying Free-Shape Masks to VLLM for Flexible Image Editing,

    T. Yuan, Y . Lin, J. Wang, Z.-Q. Cheng, X. Wang, J. GH, W. Chen, and X. Peng, “FlexEdit: Marrying Free-Shape Masks to VLLM for Flexible Image Editing,” 2025, arXiv: 2408.12429. [Online]. Available: https://arxiv.org/abs/2408.12429

  14. [22]

    Inpaint Anything: Segment Anything Meets Image Inpainting,

    T. Yu, R. Feng, R. Feng, J. Liu, X. Jin, W. Zeng, and Z. Chen, “Inpaint Anything: Segment Anything Meets Image Inpainting,” 2023, arXiv: 2304.06790. [Online]. Available: https://arxiv.org/abs/2304.06790

  15. [23]

    GLIGEN: Open-Set Grounded Text-to-Image Generation,

    Y . Li, H. Liu, Q. Wu, F. Mu, J. Yang, J. Gao, C. Li, and Y . J. Lee, “GLIGEN: Open-Set Grounded Text-to-Image Generation,” inCVPR, Jun. 2023, pp. 22 511–22 521

  16. [24]

    Adding Conditional Control to Text-to-Image Diffusion Models,

    L. Zhang, A. Rao, and M. Agrawala, “Adding Conditional Control to Text-to-Image Diffusion Models,” inICCV, Oct. 2023, pp. 3836–3847

  17. [25]

    InstructPix2Pix: Learning To Follow Image Editing Instructions,

    T. Brooks, A. Holynski, and A. A. Efros, “InstructPix2Pix: Learning To Follow Image Editing Instructions,” inCVPR, Jun. 2023, pp. 18 392– 18 402

  18. [26]

    Improving Diffusion Models for Authentic Virtual Try-on in the Wild,

    Y . Choi, S. Kwak, K. Lee, H. Choi, and J. Shin, “Improving Diffusion Models for Authentic Virtual Try-on in the Wild,” inEuropean Confer- ence on Computer Vision, 2024, pp. 206–235

  19. [27]

    CatV2TON: Taming Diffusion Transformers for Vision-Based Virtual Try-On with Temporal Concatenation,

    Z. Chong, W. Zhang, S. Zhang, J. Zheng, X. Dong, H. Li, Y . Wu, D. Jiang, and X. Liang, “CatV2TON: Taming Diffusion Transformers for Vision-Based Virtual Try-On with Temporal Concatenation,” 2025, arXiv: 2501.11325. [Online]. Available: https://arxiv.org/abs/2501.11325

  20. [28]

    CatVTON: Concatenation Is All You Need for Virtual Try-On with Diffusion Models,

    Z. Chong, X. Dong, H. Li, S. Zhang, W. Zhang, X. Zhang, H. Zhao, D. Jiang, and X. Liang, “CatVTON: Concatenation Is All You Need for Virtual Try-On with Diffusion Models,” inInternational Conference on Learning Representations, 2025

  21. [29]

    SyncDreamer: Generating Multiview-consistent Images from a Single- view Image,

    Y . Liu, C. Lin, Z. Zeng, X. Long, L. Liu, T. Komura, and W. Wang, “SyncDreamer: Generating Multiview-consistent Images from a Single- view Image,” inInternational Conference on Learning Representations, 2024

  22. [30]

    Mixed Diffusion for 3D Indoor Scene Synthesis,

    S. Hu, D. M. Arroyo, S. Debats, F. Manhardt, L. Carlone, and F. Tombari, “Mixed Diffusion for 3D Indoor Scene Synthesis,” 2024, arXiv: 2405.21066. [Online]. Available: https://arxiv.org/abs/2405.21066

  23. [31]

    SemLayoutDiff: Semantic Layout Generation with Diffusion Model for Indoor Scene Synthesis,

    X. Sun, D. Goel, and A. X. Chang, “SemLayoutDiff: Semantic Layout Generation with Diffusion Model for Indoor Scene Synthesis,” 2025, arXiv: 2508.18597. [Online]. Available: https://arxiv.org/abs/2508.18597

  24. [32]

    Scan2CAD: Learning CAD Model Alignment in RGB-D Scans,

    A. Avetisyan, M. Dahnert, A. Dai, M. Savva, A. X. Chang, and M. Niessner, “Scan2CAD: Learning CAD Model Alignment in RGB-D Scans,” inCVPR, Jun. 2019, pp. 2614–2623

  25. [33]

    PoseCNN: A Convolutional Neural Network for 6D Object Pose Estimation in Cluttered Scenes,

    Y . Xiang, T. Schmidt, V . Narayanan, and D. Fox, “PoseCNN: A Convolutional Neural Network for 6D Object Pose Estimation in Cluttered Scenes,” inRobotics: Science and Systems, 2018. [Online]. Available: https://arxiv.org/abs/1711.00199

  26. [34]

    DenseFusion: 6D Object Pose Estimation by Iterative Dense Fusion,

    C. Wang, D. Xu, Y . Zhu, R. Martin-Martin, C. Lu, L. Fei-Fei, and S. Savarese, “DenseFusion: 6D Object Pose Estimation by Iterative Dense Fusion,” inCVPR, Jun. 2019, pp. 3343–3352

  27. [35]

    Deep Hough V oting for 3D Object Detection in Point Clouds,

    C. R. Qi, O. Litany, K. He, and L. J. Guibas, “Deep Hough V oting for 3D Object Detection in Point Clouds,” inICCV, Oct. 2019, pp. 9277–9286

  28. [36]

    3D-SIS: 3D Semantic Instance Segmentation of RGB-D Scans,

    J. Hou, A. Dai, and M. Niessner, “3D-SIS: 3D Semantic Instance Segmentation of RGB-D Scans,” inCVPR, Jun. 2019, pp. 4416–4425

  29. [37]

    Point Transformer V3: Simpler, Faster, Stronger,

    X. Wu, L. Jiang, P.-S. Wang, Z. Liu, X. Liu, Y . Qiao, W. Ouyang, T. He, and H. Zhao, “Point Transformer V3: Simpler, Faster, Stronger,” inCVPR, 2024, pp. 4840–4851

  30. [38]

    3D-FRONT: 3D Furnished Rooms with layOuts and semaNTics,

    H. Fu, B. Cai, L. Gao, L. Zhang, J. Wang, C. Li, Q. Zeng, C. Sun, R. Jia, B. Zhao, and H. Zhang, “3D-FRONT: 3D Furnished Rooms with layOuts and semaNTics,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 10 913–10 922. [Online]. Available: ...

  31. [39]

    3D-FUTURE: 3D Furniture shape with TextURE,

    H. Fu, R. Jia, L. Gao, M. Gong, B. Zhao, S. Maybank, and D. Tao, “3D-FUTURE: 3D Furniture shape with TextURE,”International Journal of Computer Vision, vol. 129, pp. 3313–3337, 2021. [Online]. Available: https://arxiv.org/abs/2009.09633

  32. [40]

    BlenderProc2: A Procedural Pipeline for Photorealistic Rendering,

    M. Denninger, D. Winkelbauer, M. Sundermeyer, W. Boerdijk, M. Knauer, K. H. Strobl, M. Humt, and R. Triebel, “BlenderProc2: A Procedural Pipeline for Photorealistic Rendering,”Journal of Open Source Software, vol. 8, no. 82, p. 4901, 2023

  33. [41]

    Grounded SAM: Assembling Open-World Models for Diverse Visual Tasks,

    T. Ren, S. Liu, A. Zeng, J. Lin, K. Li, H. Cao, J. Chen, X. Huang, Y . Chen, F. Yan, Z. Zeng, H. Zhang, F. Li, J. Yang, H. Li, Q. Jiang, and L. Zhang, “Grounded SAM: Assembling Open-World Models for Diverse Visual Tasks,” 2024, arXiv: 2401.14159. [Online]. Available: https://a...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.