pith. sign in

arxiv: 2604.10675 · v1 · submitted 2026-04-12 · 💻 cs.CV

HiddenObjects: Scalable Diffusion-Distilled Spatial Priors for Object Placement

Marco Schouten , Ioannis Siglidis , Serge Belongie , Dim P. Papadopoulos This is my paper

Pith reviewed 2026-05-10 16:04 UTC · model grok-4.3

classification 💻 cs.CV
keywords object placementspatial priorsdiffusion modelsinpaintingimage editingdatasetcomputer vision
0
0 comments X

The pith

Spatial priors for placing objects in scenes can be learned at scale by distilling implicit knowledge from text-conditioned diffusion models, yielding better results than human annotations.

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

The paper develops an automated pipeline that uses diffusion-based inpainting to score and rank millions of possible object placements across real background scenes, thereby extracting explicit class-conditioned spatial priors without manual labeling. This process builds a dataset of 27 million annotations that trains models to insert objects more naturally than sparse human data or existing baselines. The approach matters because realistic placement determines the quality of image editing and synthesis tasks, and scaling beyond limited annotations could support broader applications in computer vision. The work further compresses the priors into a lightweight model that runs far faster for practical use.

Core claim

Implicit knowledge of natural object-scene relationships encoded in diffusion models can be made explicit by densely evaluating placement candidates through inpainting on authentic backgrounds, producing spatial priors that outperform both human-annotated placements and prior methods on image editing benchmarks.

What carries the argument

The diffusion-based inpainting pipeline that evaluates dense bounding-box insertions on real scenes to generate ranked, class-conditioned placement scores.

If this is right

  • The distilled priors improve object insertion quality in downstream image editing over sparse human annotations.
  • A lightweight model distilled from the priors runs 230000 times faster than the original evaluation pipeline.
  • The priors surpass both existing placement algorithms and zero-shot vision-language models on object placement tasks.
  • The framework enables construction of large-scale placement datasets across many scenes and categories without manual effort.

Where Pith is reading between the lines

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

  • The same distillation process could be repeated with newer diffusion models to iteratively refine the priors.
  • These priors might reduce unnatural artifacts when used to guide other generative image models.
  • Extending the pipeline to video or 3D scenes could produce temporal or depth-aware placement models.

Load-bearing premise

The inpainting scores measure genuine placement naturalness rather than diffusion model artifacts or biases that might not transfer to editing tasks.

What would settle it

Human evaluators consistently rating high-scoring placements from the pipeline as less realistic than low-scoring ones on a held-out set of scenes.

Figures

Figures reproduced from arXiv: 2604.10675 by Dim P. Papadopoulos, Ioannis Siglidis, Marco Schouten, Serge Belongie.

Figure 1
Figure 1. Figure 1: Spatial Priors for Object Placements. (a) We score candidate bounding boxes for object insertion using a diffusion-based inpainting pipeline. In this exam￾ple, three pizza locations are considered. First, we attempt to insert the object in the designated locations via inpainting. Then, we evaluate these locations, yielding three possible outcomes: (i) “Pizza Detected”, (ii) “Wrong Object” or (iii) “No Obje… view at source ↗
Figure 2
Figure 2. Figure 2: Background artifacts in PIPE [52]. Object removal pipelines are effective at removing the requested object but leave visible traces that can facilitate shortcut learning, e.g., (left) blurry area and the shadow presence, or (right) larger artifacts. across verified placements. Exhaustive testing of multiple insertions using a slid￾ing window can densely capture and rank all plausible object locations withi… view at source ↗
Figure 3
Figure 3. Figure 3: Spatial Prior Extraction. Given a background image x, the inpainter I synthesizes an image ˆxij for target class cj at a bounding box bi. The verifier V evaluates the inpainted images ˆxij , yielding a detected bounding box and a human￾aligned preference score rij . Repeating this process across multiple sampled boxes yields the ground-truth spatial prior S(x, cj ) for the j-th class. The heatmap visualize… view at source ↗
Figure 4
Figure 4. Figure 4: Spatial Prior (HiddenObjects) aggregated as a heatmap. (Left) We visualize the mean aggregated spatial prior across all images of HiddenObjects for different classes listed on the left. Classes such as bench and boat tend to appear near the bottom of images, while kite occurs more frequently near the top. These patterns reflect both semantic regularities (e.g., kite in the sky) and inherent photographic bi… view at source ↗
Figure 5
Figure 5. Figure 5: Verifier Comparison. For each verifier, we display the inpainting generation of the highest-scoring location. CLIP Score often prefers insertions irrespective of global scene context. In contrast, Image and Vision Reward outperform aesthetic score in judging realism and semantic coherence, leading to more human-aligned spatial priors. Speedup. For the majority of insertion proposals, we expect that the inp… view at source ↗
Figure 6
Figure 6. Figure 6: Inpainter Comparison. We compare cross-attention-based inpainters with ControlNet-based variants. Cross-attention-based inpainters are prone to (i) inpainting artifacts or (ii) inpaint objects that don’t respect scene semantics. While ControlNet￾based methods more reliably enforce scene constraints: they either (iii) refuse implau￾sible insertions or generate (valid) semantically coherent, context-aware ob… view at source ↗
Figure 7
Figure 7. Figure 7: Image inpainting with object placement priors. [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Analysis of Spatial and Scale Distributions. (a) [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Image inpainting with object placement priors. [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Composite Overlay of Inpaintings. Multiple valid inpaintings are aggre￾gated into a single composite scene. Accepted objects are segmented and overlaid on the source background. Insertions are ordered by preference ranking so the highest￾ranked object remains unoccluded at the top of the stack. B Spatial Prior Visualization Building upon the aggregated spatial prior distributions introduced in [PITH_FULL… view at source ↗
Figure 11
Figure 11. Figure 11: Placement Distillation Model. Given a background image and a target object class, the model predicts plausible insertion locations using a DETR-style ar￾chitecture. The transformer decoder operates on a fixed set of learned queries that are conditioned on the target class embedding. Each query predicts a candidate bounding box with its plausibility score. During inference, predicted boxes are ranked by th… view at source ↗
Figure 12
Figure 12. Figure 12: Speedup Analysis. Top: Distribution of ∆(xt, ∅) for 10040 inpaintings. Bottom: Plotting recall versus time across time steps shows that using only the second denoising iteration seems to be sufficient to get a 2x speed-up improvement on an 80% recall. The dotted red line denotes the cost of skipping the first step all along N · τ20. unconditional branch c = ∅. Our observation is simple: we measure how muc… view at source ↗
Figure 13
Figure 13. Figure 13: Bounding Box Proposals. Left: The 12 × 12 anchor grid. Middle: Full coverage set for an interior anchor. Right: Pruned set for a corner anchor due to boundary constraints. across timestep. We notice that there is a fine separation between successful and unsuccessful generations, where successful ones exhibit much higher values (mean 1.47) than failed ones (mean 0.38). While in step-5 generations are bette… view at source ↗
Figure 14
Figure 14. Figure 14: Spatial Prior. Object placements are computed for the following classes: airplane, apple, bench, bicycle, boat, book, bottle, cake [PITH_FULL_IMAGE:figures/full_fig_p026_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Spatial Prior. Object placements are computed for the following classes: car, cat, chair, cow, cup, dog, elephant, fire hydrant [PITH_FULL_IMAGE:figures/full_fig_p027_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Spatial Prior. Object placements are computed for the following classes: horse, keyboard, laptop, motorcycle, person, pizza, potted plant, sandwich [PITH_FULL_IMAGE:figures/full_fig_p028_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Inpainting with Object Placement Priors. [PITH_FULL_IMAGE:figures/full_fig_p029_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Inpainting with Object Placement Priors. [PITH_FULL_IMAGE:figures/full_fig_p030_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Inpainting with Object Placement Priors. [PITH_FULL_IMAGE:figures/full_fig_p031_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Inpainting with Object Placement Priors. [PITH_FULL_IMAGE:figures/full_fig_p032_20.png] view at source ↗
read the original abstract

We propose a method to learn explicit, class-conditioned spatial priors for object placement in natural scenes by distilling the implicit placement knowledge encoded in text-conditioned diffusion models. Prior work relies either on manually annotated data, which is inherently limited in scale, or on inpainting-based object-removal pipelines, whose artifacts promote shortcut learning. To address these limitations, we introduce a fully automated and scalable framework that evaluates dense object placements on high-quality real backgrounds using a diffusion-based inpainting pipeline. With this pipeline, we construct HiddenObjects, a large-scale dataset comprising 27M placement annotations, evaluated across 27k distinct scenes, with ranked bounding box insertions for different images and object categories. Experimental results show that our spatial priors outperform sparse human annotations on a downstream image editing task (3.90 vs. 2.68 VLM-Judge), and significantly surpass existing placement baselines and zero-shot Vision-Language Models for object placement. Furthermore, we distill these priors into a lightweight model for fast practical inference (230,000x faster).

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes HiddenObjects, a dataset of 27M ranked placement annotations generated by applying a diffusion-based inpainting pipeline to evaluate dense object insertions on 27k real backgrounds. Spatial priors are learned from this data and distilled into a lightweight model; the central claim is that these priors outperform sparse human annotations (3.90 vs. 2.68 VLM-Judge) on a downstream image-editing task, exceed existing placement baselines and zero-shot VLMs, and enable 230,000x faster inference.

Significance. If the evaluation is shown to be independent of the diffusion model and VLM-Judge biases, the work would provide a scalable, annotation-free route to explicit spatial priors that meaningfully improves object placement in editing pipelines. The reported inference speedup would also make the approach practically deployable.

major comments (3)
  1. [Abstract] Abstract: The headline result (3.90 vs. 2.68 VLM-Judge on the editing task) is presented without any description of the inpainting pipeline controls, the exact ranking procedure for the 27M placements, statistical significance tests, or validation that the VLM-Judge correlates with human naturalness judgments. These omissions make it impossible to assess whether the reported gains reflect genuine placement quality or pipeline artifacts.
  2. [Evaluation] Evaluation / downstream task: The manuscript does not report any ablation that removes or replaces the diffusion-inpainting component, nor any correlation analysis between the VLM-Judge and the diffusion model family used to generate the training scores. Without such checks, the superiority claim risks circularity, as both the training signal and the evaluation metric could encode the same model biases rather than independent naturalness.
  3. [§4] §4 (or equivalent experiments section): No human-study validation or cross-model judge comparison is provided to confirm that the VLM-Judge metric is orthogonal to the diffusion model used for data generation. This is load-bearing for the central claim that the distilled priors transfer to real editing tasks without shortcut learning.
minor comments (2)
  1. [Dataset construction] The description of scene and category diversity in the 27k backgrounds and 27M annotations would benefit from explicit statistics (e.g., object category distribution, scene type coverage) to allow readers to judge generalization.
  2. [Distillation] Notation for the distilled lightweight model (architecture, training objective, exact speedup measurement) is introduced only in the abstract and would be clearer with a dedicated paragraph or table in the main text.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the insightful comments on our paper. We have carefully considered each point and provide detailed responses below. Where appropriate, we will revise the manuscript to address the concerns raised, particularly by adding more methodological details, ablations, and validation studies.

read point-by-point responses

  1. Referee: The headline result (3.90 vs. 2.68 VLM-Judge on the editing task) is presented without any description of the inpainting pipeline controls, the exact ranking procedure for the 27M placements, statistical significance tests, or validation that the VLM-Judge correlates with human naturalness judgments. These omissions make it impossible to assess whether the reported gains reflect genuine placement quality or pipeline artifacts.

    Authors: We agree that additional details in the abstract would improve clarity. In the revised manuscript, we will expand the abstract to include a brief overview of the inpainting pipeline controls and the ranking procedure used to generate the 27M annotations. We will also report p-values or statistical significance for the VLM-Judge comparisons. For the correlation of VLM-Judge with human judgments, we will add a new analysis in the supplementary material showing agreement on a held-out set of images. revision: yes

  2. Referee: The manuscript does not report any ablation that removes or replaces the diffusion-inpainting component, nor any correlation analysis between the VLM-Judge and the diffusion model family used to generate the training scores. Without such checks, the superiority claim risks circularity, as both the training signal and the evaluation metric could encode the same model biases rather than independent naturalness.

    Authors: We acknowledge the potential for circularity and the value of such ablations. While the training signal comes from diffusion-based inpainting scores and the evaluation uses a separate VLM-Judge, we will add in the revision a correlation analysis between the diffusion scores and VLM-Judge scores on the same placements to demonstrate they capture distinct aspects. We will also include an ablation study replacing the diffusion inpainting with a non-diffusion baseline where possible, and discuss how using real backgrounds and dense placements mitigates shortcut learning. This will be detailed in the experiments section. revision: yes

  3. Referee: No human-study validation or cross-model judge comparison is provided to confirm that the VLM-Judge metric is orthogonal to the diffusion model used for data generation. This is load-bearing for the central claim that the distilled priors transfer to real editing tasks without shortcut learning.

    Authors: We agree that additional validation would strengthen the work. In the revised manuscript, we will include results from a small-scale human study on a subset of the image editing outputs to correlate VLM-Judge scores with human naturalness ratings. We will also perform and report cross-model judge comparisons using an alternative vision-language model to verify consistency and orthogonality to the original diffusion model family. These will be added to Section 4. revision: yes

Circularity Check

0 steps flagged

No significant circularity: derivation is data-driven and externally evaluated.

full rationale

The paper generates a large placement dataset (27M annotations) via diffusion inpainting on real backgrounds, trains class-conditioned spatial priors on the resulting ranked scores, and evaluates the priors on a downstream image-editing task using an independent VLM-Judge metric. No equations, self-definitions, or self-citations are presented that reduce any load-bearing claim (e.g., the 3.90 vs 2.68 VLM-Judge result or outperformance of baselines) to the input scores by construction. The pipeline is explicitly described as distilling implicit knowledge into explicit priors and then testing transfer, with no renaming of known results or ansatz smuggling. The derivation chain therefore remains self-contained against the stated external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the unproven premise that text-conditioned diffusion models encode accurate, transferable spatial placement knowledge and that an inpainting-based evaluation pipeline can extract it without systematic bias. No explicit free parameters or new invented entities are named in the abstract.

axioms (1)
  • domain assumption Text-conditioned diffusion models contain implicit, class-conditioned knowledge about realistic object placement in natural scenes.
    The entire distillation approach depends on this assumption being true and extractable via inpainting.

pith-pipeline@v0.9.0 · 5488 in / 1388 out tokens · 58914 ms · 2026-05-10T16:04:11.126738+00:00 · methodology

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

68 extracted references · 68 canonical work pages · 5 internal anchors

  1. [1]

    In: ICCV (2025)

    Abdelreheem, A., Aleotti, F., Watson, J., Qureshi, Z., Eldesokey, A., Wonka, P., Brostow, G., Vicente, S., Garcia-Hernando, G.: Placeit3d: Language-guided object placement in real 3d scenes. In: ICCV (2025)

    work page 2025

  2. [2]

    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

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  3. [3]

    FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space

    Batifol, S., Blattmann, A., Boesel, F., Consul, S., Diagne, C., Dockhorn, T., En- glish, J., English, Z., Esser, P., Kulal, S., et al.: Flux. 1 kontext: Flow matching for in-context image generation and editing in latent space. arxivabs/2506.15742 (2025)

    work page internal anchor Pith review arXiv 2025

  4. [4]

    a is b” fail to learn “b is a

    Bau,D.,Zhu,J.Y.,Strobelt,H.,Lapedriza,A.,Zhou,B.,Torralba,A.:Understand- ing the role of individual units in a deep neural network. Proceedings of the Na- tional Academy of Sciences (2020).https://doi.org/10.1073/pnas.1907375117, https://www.pnas.org/content/early/2020/08/31/1907375117

    work page doi:10.1073/pnas.1907375117 2020

  5. [5]

    In: ICLR (2019)

    Bau, D., Zhu, J.Y., Strobelt, H., Zhou, B., Tenenbaum, J.B., Freeman, W.T., Torralba, A.: Gan dissection: Visualizing and understanding generative adversarial networks. In: ICLR (2019)

    work page 2019

  6. [6]

    ACM Comput

    Boukerche, A., Zheng, L., Alfandi, O.: Outlier detection: Methods, models, and classification. ACM Comput. Surv.53(3), 55:1–55:37 (2021)

    work page 2021

  7. [7]

    In: ECCV

    Canet Tarrés, G., Lin, Z., Zhang, Z., Zhang, J., Song, Y., Ruta, D., Gilbert, A., Collomosse, J., Kim, S.Y.: Thinking outside the bbox: Unconstrained generative object compositing. In: ECCV. pp. 476–495 (2024)

    work page 2024

  8. [8]

    Referring layer decomposition.arXiv preprint arXiv:2602.19358, 2026

    Chen, F., Shen, Y., Xu, L., Yuan, Y., Zhang, S., Niu, Y., Wen, L.: Referring layer decomposition. arXiv preprint arXiv:2602.19358 (2026)

    work page arXiv 2026

  9. [9]

    In: CVPR

    Chen, Z., Wu, J., Wang, W., Su, W., Chen, G., Xing, S., Zhong, M., Zhang, Q., Zhu, X., Lu, L., et al.: Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In: CVPR. pp. 24185–24198 (2024)

    work page 2024

  10. [10]

    Neurocomputing p

    Cheng, X., Zhai, P., Yang, D., Meng, X., Xia, Y., Zhang, L.: Diverse object place- ment with dual interaction. Neurocomputing p. 131161 (2025)

    work page 2025

  11. [11]

    , author Dong, W

    Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large- scale hierarchical image database. In: CVPR. pp. 248–255 (2009).https://doi. org/10.1109/CVPR.2009.5206848

    work page doi:10.1109/cvpr.2009.5206848 2009

  12. [12]

    In: ICCV

    Doersch, C., Gupta, A., Efros, A.A.: Unsupervised visual representation learning by context prediction. In: ICCV. pp. 1422–1430 (2015)

    work page 2015

  13. [13]

    In: ECCV

    Dvornik, N., Mairal, J., Schmid, C.: Modeling visual context is key to augmenting object detection datasets. In: ECCV. pp. 364–380 (2018)

    work page 2018

  14. [14]

    http://www.pascal-network.org/challenges/VOC/voc2012/workshop/index.html 16 M

    Everingham, M., Van Gool, L., Williams, C.K.I., Winn, J., Zisserman, A.: The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascal-network.org/challenges/VOC/voc2012/workshop/index.html 16 M. Schouten et al

    work page 2012

  15. [15]

    In: CVPR (2008)

    Galleguillos, C., Rabinovich, A., Belongie, S.: Object categorization using co- occurrence, location and appearance. In: CVPR (2008)

    work page 2008

  16. [16]

    ICME (2025)

    Gao, B., Zhang, B., Niu, L.: Object placement for anything. ICME (2025)

    work page 2025

  17. [17]

    In: ICCV (2025)

    Gao, J., Joseph, K.J., la Torre, F.D.: Teleportraits: Training-free people insertion into any scene. In: ICCV (2025)

    work page 2025

  18. [18]

    In: CVPR (2019)

    Gupta, A., Dollar, P., Girshick, R.: LVIS: A dataset for large vocabulary instance segmentation. In: CVPR (2019)

    work page 2019

  19. [19]

    ACM Trans- actions on graphics (TOG)26(3), 4–es (2007)

    Hays, J., Efros, A.A.: Scene completion using millions of photographs. ACM Trans- actions on graphics (TOG)26(3), 4–es (2007)

    work page 2007

  20. [20]

    arXiv preprint arXiv:2412.14462 (2024)

    He, J., Li, W., Liu, Y., Kim, J., Wei, D., Pfister, H.: Affordance-aware object insertion via mask-aware dual diffusion. arXiv preprint arXiv:2412.14462 (2024)

    work page arXiv 2024

  21. [21]

    In: Moens, M., Huang, X., Specia, L., Yih, S.W

    Hessel, J., Holtzman, A., Forbes, M., Bras, R.L., Choi, Y.: Clipscore: A reference- free evaluation metric for image captioning. In: Moens, M., Huang, X., Specia, L., Yih, S.W. (eds.) EMNLP. pp. 7514–7528. Association for Computational Linguis- tics (2021)

    work page 2021

  22. [22]

    In: CVPR (2025)

    Huang, I., Bao, Y., Truong, K., Zhou, H., Schmid, C., Guibas, L., Fathi, A.: Fire- place: Geometric refinements of llm common sense reasoning for 3d object place- ment. In: CVPR (2025)

    work page 2025

  23. [23]

    In: CVPR

    Kulal, S., Brooks, T., Aiken, A., Wu, J., Yang, J., Lu, J., Efros, A.A., Singh, K.K.: Putting people in their place: Affordance-aware human insertion into scenes. In: CVPR. pp. 17089–17099 (2023)

    work page 2023

  24. [24]

    NeurIPS31(2018)

    Lee, D., Liu, S., Gu, J., Liu, M.Y., Yang, M.H., Kautz, J.: Context-aware synthesis and placement of object instances. NeurIPS31(2018)

    work page 2018

  25. [25]

    ACM Multimedia (2025)

    Li, C., Wang, W., Li, Q., Lepri, B., Sebe, N., Nie, W.: Freeinsert: Disentangled text-guided object insertion in 3d gaussian scene without spatial priors. ACM Multimedia (2025)

    work page 2025

  26. [26]

    Li, T., Ku, M., Wei, C., Chen, W.: Dreamedit: Subject-driven image editing. Trans. Mach. Learn. Res. (2023)

    work page 2023

  27. [27]

    arXiv preprint arXiv:2507.16813 (2025)

    Liang, D., Jia, J., Liu, Y., Lau, R.W.: Hocomp: Interaction-aware human-object composition. arXiv preprint arXiv:2507.16813 (2025)

    work page arXiv 2025

  28. [28]

    Lin, C.H., Yumer, E., Wang, O., Shechtman, E., Lucey, S.: St-gan: Spatial trans- formergenerativeadversarialnetworksforimagecompositing.In:CVPR.pp.9455– 9464 (2018)

    work page 2018

  29. [29]

    In: ECCV

    Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: Common objects in context. In: ECCV. pp. 740–755 (2014)

    work page 2014

  30. [30]

    In: CVPR

    Liu, H., Li, C., Li, Y., Lee, Y.J.: Improved baselines with visual instruction tuning. In: CVPR. pp. 26296–26306 (2024)

    work page 2024

  31. [31]

    arXiv preprint arXiv:2107.01889 (2021)

    Liu, L., Liu, Z., Zhang, B., Li, J., Niu, L., Liu, Q., Zhang, L.: Opa: object placement assessment dataset. arXiv preprint arXiv:2107.01889 (2021)

    work page arXiv 2021

  32. [32]

    arXiv preprint arXiv:2309.15508 (2023)

    Lu, L., Li, J., Zhang, B., Niu, L.: Dreamcom: Finetuning text-guided inpainting model for image composition. arXiv preprint arXiv:2309.15508 (2023)

    work page arXiv 2023

  33. [33]

    arXiv preprint arXiv:2205.14280 (2022)

    Niu, L., Liu, Q., Liu, Z., Li, J.: Fast object placement assessment. arXiv preprint arXiv:2205.14280 (2022)

    work page arXiv 2022

  34. [34]

    In: CVPR (2019)

    Papadopoulos, D.P., Tamaazousti, Y., Ofli, F., Weber, I., Torralba, A.: How to make a pizza: Learning a compositional layer-based gan model. In: CVPR (2019)

    work page 2019

  35. [35]

    In: ECCV

    Parihar, R., Gupta, H., VS, S., Babu, R.V.: Text2place: Affordance-aware text guided human placement. In: ECCV. pp. 57–77 (2024)

    work page 2024

  36. [36]

    In: CVPR (2025) HiddenObjects17

    Parihar, R., Sarkar, S., Vora, S., Kundu, J., Babu, R.V.: Monoplace3d: Learning 3d-aware object placement for 3d monocular detection. In: CVPR (2025) HiddenObjects17

    work page 2025

  37. [37]

    arXiv preprint arXiv:2504.17076 (2025)

    Petersen, J., Abati, D., Habibian, A., Wiggers, A.: Scene-aware location mod- eling for data augmentation in automotive object detection. arXiv preprint arXiv:2504.17076 (2025)

    work page arXiv 2025

  38. [38]

    In: ICLR (2024)

    Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., Müller, J., Penna, J., Rombach, R.: Sdxl: Improving latent diffusion models for high-resolution image synthesis. In: ICLR (2024)

    work page 2024

  39. [39]

    In: CVPR

    Poska, M., Huang, S.X., Hwang, B.: Hopnet: Harmonizing object placement net- work for realistic image generation via object composition. In: CVPR. pp. 6344– 6354 (2025)

    work page 2025

  40. [40]

    In: ECCV

    Qin, Y., Xu, J., Wang, R., Chen, X.: Think before placement: Common sense enhanced transformer for object placement. In: ECCV. pp. 35–50 (2024)

    work page 2024

  41. [41]

    In: ICCV (2007)

    Rabinovich, A., Vedaldi, A., Galleguillos, C., Wiewiora, E., Belongie, S.: Objects in context. In: ICCV (2007)

    work page 2007

  42. [42]

    Hierarchical Text-Conditional Image Generation with CLIP Latents

    Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., Chen, M.: Hierarchical text- conditional image generation with clip latents. arXiv preprint arXiv:2204.06125 1(2), 3 (2022)

    work page internal anchor Pith review arXiv 2022

  43. [43]

    Ren, T., Liu, S., Zeng, A., Lin, J., Li, K., Cao, H., Chen, J., Huang, X., Chen, Y., Yan, F., Zeng, Z., Zhang, H., Li, F., Yang, J., Li, H., Jiang, Q., Zhang, L.: Grounded sam: Assembling open-world models for diverse visual tasks (2024)

    work page 2024

  44. [44]

    In: CVPR (2021)

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2021)

    work page 2021

  45. [45]

    In: SCIA (2025)

    Schouten, M., Kaya, M.O., Belongie, S., Papadopoulos, D.P.: Poem: Precise object- level editing via mllm control. In: SCIA (2025)

    work page 2025

  46. [46]

    Schuhmann, C.: Aesthetic predictor v2.5.https://github.com/discus0434/ aesthetic-predictor-v2-5/(may 2024), gitHub repository

    work page 2024

  47. [47]

    In: NeurIPS (2022)

    Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: Laion-5b: An open large-scale dataset for training next generation image-text models. In: NeurIPS (2022)

    work page 2022

  48. [48]

    In: ICLR (2025), https://openreview.net/forum?id=ZeaTvXw080

    Tewel,Y.,Gal,R.,Samuel,D.,Atzmon,Y.,Wolf,L.,Chechik,G.:Add-it:Training- free object insertion in images with pretrained diffusion models. In: ICLR (2025), https://openreview.net/forum?id=ZeaTvXw080

    work page 2025

  49. [49]

    In: ICCV

    Torralba: Context-based vision system for place and object recognition. In: ICCV. pp. 273–280 (2003)

    work page 2003

  50. [50]

    In: CVPR

    Tripathi, S., Chandra, S., Agrawal, A., Tyagi, A., Rehg, J.M., Chari, V.: Learning to generate synthetic data via compositing. In: CVPR. pp. 461–470 (2019)

    work page 2019

  51. [51]

    In: ECCV Workshop

    Volokitin, A., Susmelj, I., Agustsson, E., Van Gool, L., Timofte, R.: Efficiently detecting plausible locations for object placement using masked convolutions. In: ECCV Workshop. pp. 252–266 (2020)

    work page 2020

  52. [52]

    In: CVPR (2025)

    Wasserman, N., Rotstein, N., Ganz, R., Kimmel, R.: Paint by inpaint: Learning to add image objects by removing them first. In: CVPR (2025)

    work page 2025

  53. [53]

    In: ICCV

    Winter,D.,Shul,A.,Cohen,M.,Berman,D.,Pritch,Y.,Rav-Acha,A.,Hoshen,Y.: Objectmate: A recurrence prior for object insertion and subject-driven generation. In: ICCV. pp. 16281–16291 (2025)

    work page 2025

  54. [54]

    Qwen-Image Technical Report

    Wu, C., Li, J., Zhou, J., Lin, J., Gao, K., Yan, K., Yin, S.m., Bai, S., Xu, X., Chen, Y., et al.: Qwen-image technical report. arXiv preprint arXiv:2508.02324 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  55. [55]

    In: AAAI

    Xu, J., Huang, Y., Cheng, J., Yang, Y., Xu, J., Wang, Y., Duan, W., Yang, S., Jin, Q., Li, S., Teng, J., Yang, Z., Zheng, W., Liu, X., Zhang, D., Ding, M., Zhang, X., Huang, S., Gu, X., Huang, M., Tang, J., Dong, Y.: Visionreward: Fine-grained multi-dimensional human preference learning for image and video generation. In: AAAI. pp. 11269–11277 (2026) 18 M...

    work page 2026

  56. [56]

    In: NeurIPS

    Xu, J., Liu, X., Wu, Y., Tong, Y., Li, Q., Ding, M., Tang, J., Dong, Y.: Imagere- ward: learning and evaluating human preferences for text-to-image generation. In: NeurIPS. pp. 15903–15935 (2023)

    work page 2023

  57. [57]

    ImgEdit: A Unified Image Editing Dataset and Benchmark

    Ye, Y., He, X., Li, Z., Lin, B., Yuan, S., Yan, Z., Hou, B., Yuan, L.: Imgedit: A unified image editing dataset and benchmark. arXiv preprint arXiv:2505.20275 (2025)

    work page internal anchor Pith review arXiv 2025

  58. [58]

    In: NeurIPS (2023)

    Yuan, L., Hong, J., Sarukkai, V., Fatahalian, K.: Learning to place objects into scenes by hallucinating scenes around objects. In: NeurIPS (2023)

    work page 2023

  59. [59]

    Yun, J., Abati, D., Omran, M., Choo, J., Habibian, A., Wiggers, A.: Imagining the unseen: Generative location modeling for object placement (2025),https: //arxiv.org/abs/2410.13564

    work page arXiv 2025

  60. [60]

    In: ECCV

    Zhang, L., Wen, T., Min, J., Wang, J., Han, D., Shi, J.: Learning object placement by inpainting for compositional data augmentation. In: ECCV. pp. 566–581 (2020)

    work page 2020

  61. [61]

    In: CVPR

    Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: CVPR. pp. 3836–3847 (2023)

    work page 2023

  62. [62]

    Computational Visual Media 6(1), 79–93 (2020)

    Zhang, S.H., Zhou, Z.P., Liu, B., Dong, X., Hall, P.: What and where: A context- based recommendation system for object insertion. Computational Visual Media 6(1), 79–93 (2020)

    work page 2020

  63. [63]

    IEEE transactions on pattern analysis and machine intelligence40(6), 1452–1464 (2017)

    Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence40(6), 1452–1464 (2017)

    work page 2017

  64. [64]

    IJCV127(3), 302– 321 (2019)

    Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semantic understanding of scenes through the ade20k dataset. IJCV127(3), 302– 321 (2019)

    work page 2019

  65. [65]

    IEEE Transactions on Visualization and Com- puter Graphics30(7), 3151–3165 (2022)

    Zhou, H., Ma, R., Zhang, L.X., Gao, L., Mahdavi-Amiri, A., Zhang, H.: Sac-gan: Structure-aware image composition. IEEE Transactions on Visualization and Com- puter Graphics30(7), 3151–3165 (2022)

    work page 2022

  66. [66]

    In: CVPR

    Zhou, H., Zuo, X., Ma, R., Cheng, L.: Bootplace: Bootstrapped object placement with detection transformers. In: CVPR. pp. 19294–19303 (2025)

    work page 2025

  67. [67]

    In: ECCV

    Zhou, S., Liu, L., Niu, L., Zhang, L.: Learning object placement via dual-path graph completion. In: ECCV. pp. 373–389 (2022)

    work page 2022

  68. [68]

    apple”, “cat

    Zhu, S., Lin, Z., Cohen, S., Kuen, J., Zhang, Z., Chen, C.: Topnet: Transformer- based object placement network for image compositing. In: CVPR. pp. 1838–1847 (2023) HiddenObjectsA-1 Appendix potted plant Input BG Human Annot. Ours elephant Input BG Human Annot. Ours keyboard car elephant cat bicycle pizza Fig.9: Image inpainting with object placement pri...

    work page 2023