Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

A two-stage model, Interact-Custom, generates images of a specified person and object interacting while preserving both identities, using a text-generated foreground mask as the spatial guide for the interaction.

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 ↗

Interact-Custom generates customized human-object interaction images by first generating a foreground mask from the prompt and then using that mask to guide identity-preserving diffusion generation.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A solid new task and a plausible two-stage method; the main caveat is the untested mask bottleneck that carries all interaction semantics. the 4 major comments →

arxiv 2508.19575 v2 pith:GS6CWGOB submitted 2025-08-27 cs.CV cs.AI

Interact-Custom: Customized Human Object Interaction Image Generation

Zhu Xu , Zhaowen Wang , Yuxin Peng , Yang Liu This is my paper
classification cs.CV cs.AI
keywords human-object interactionimage customizationidentity preservationinteraction semanticsmask-guided generationtwo-stage diffusion modeldataset constructiontext-to-image generation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper proposes a new generation task, customized human-object interaction image generation, in which a single output image must show a specified person and a specified object performing a described interaction while both look like the supplied references. It argues that existing multi-concept customization models preserve appearance but lose fine-grained interaction, because they never control the spatial layout of the humans and objects. The proposed two-stage model, Interact-Custom, first turns the text prompt into a foreground mask of the interacting pair, and then uses that mask as spatial guidance while injecting identity features of the target person and object. To learn this, the authors build a dataset of about one million samples in which the same human-object pair appears in many different interactive poses. Evaluations on identity fidelity and on interaction semantics report gains over existing customization approaches, with the largest benefits on keeping the interaction understandable.

Core claim

The central claim is that interaction semantics in customized generation can be controlled by explicitly separating 'who' from 'how': a text-conditioned diffusion model first generates a binary foreground mask of the human and object whose shape encodes contact points and relative positions; a second diffusion model then generates the actual image, using the mask both as a shape constraint and as the carrier of interaction meaning, while multi-granularity appearance features of the reference human and object are injected to preserve identity. The authors demonstrate the claim on a newly collected dataset of roughly one million samples with identical human-object pairs in different poses, des

What carries the argument

The load-bearing mechanism is the Interaction-Aware Mask, an explicit foreground mask generated from the text prompt alone, which acts as the sole spatial configuration signal passed to the image generator. Because the mask's shape and relative placement determine which body parts meet and how close the object is, it converts the abstract interaction verb into concrete geometry; the image generator then treats the mask as guidance for pose and contact while a cross-attention path carries identity features from the reference images. Coarse patch features from a self-supervised vision transformer, combined with edge-detection detail maps, carry the appearance, while the mask carries the intera

Load-bearing premise

The load-bearing premise is that a text-only generated binary mask carries enough contact and relative-position detail to specify the interaction; if the mask is wrong, the final image cannot recover the correct action, because the image generator treats the mask as the interaction's spatial truth.

What would settle it

A direct falsifier is a mask-perturbation experiment: for a contact-sensitive prompt, swap in a mask that separates the hand from the object while keeping the same action label, and check whether the final image still shows the action under a human-object interaction detector. If the detected interaction score does not drop, the mask is not the carrier of interaction semantics; if the image degrades, the paper's mechanism is confirmed.

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

If this is right

  • Users can steer an interaction by editing the text prompt, since the first stage regenerates a mask that reflects the new action.
  • The same identity can be placed into many different poses and contacts by swapping only the mask, which is the intended decomposition of identity and interaction features.
  • Providing a background image and a bounding box lets the interaction be composed into a scene, a natural requirement for advertising and illustration.
  • The reported gap to real images on spatial-sensitive metrics gives a concrete target: making mask geometry closer to ground-truth human-object interaction layouts should improve the final image.

Where Pith is reading between the lines

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

  • The mask-first recipe likely generalizes beyond humans: any pairwise interaction between two customizable objects could use the same two-stage decomposition, provided training data with matched identity pairs and varied poses exist.
  • Adding part-level contact annotations to the mask would be a direct test of whether finer spatial configuration closes the gap to real images that the paper reports.
  • The paper's retrieval-based mask bank underperforms the generation-based one by a small margin; a retriever that matches morphology as well as text would likely narrow that gap, since the paper attributes the loss to incomplete morphological correspondence.
  • The dataset built from same-pair/different-pose videos could be reused to probe feature disentanglement directly, for example by measuring whether identity features stay invariant under pose changes in the trained model.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper introduces a new task, Customized Human Object Interaction (CHOI) image generation, which requires preserving the identity of a target human and object while generating an image of them performing a specified interaction. To support this task, the authors construct a large dataset of roughly 1M samples by adapting static HOI images and mining video clips, resulting in samples of the same human-object pair in different poses. They propose a two-stage model, Interact-Custom: first, Interaction-Aware Mask Generation (IAMG) generates a binary foreground mask conditioned only on the interaction text prompt; second, Mask Guided Image Generation (MGIG), built on AnyDoor, uses the mask for spatial guidance and DINOv2/Sobel features for identity preservation to synthesize the final image. The paper reports quantitative improvements over customization baselines on CLIP/DINO identity scores, LLaVA-based holistic semantic accuracy, and spatial semantic mAP, along with a user study and ablations. The central empirical claim is that Interact-Custom outperforms existing compositional customization methods on identity fidelity and interaction semantic control.

Significance. If the claims hold, this is a useful step toward combining subject customization with fine-grained interaction control, a largely unexplored intersection. The dataset construction effort and the explicit two-stage decomposition are valuable contributions; the paper ships a code URL and a new dataset, which aids reproducibility. The identity preservation gains over strong baselines such as AnyDoor are notable and well supported by the reported numbers. However, the paper's central interaction-control claim rests on the assumption that a binary foreground mask is a sufficient carrier of interaction semantics from the text prompt to the image generator. That assumption is not directly validated, and several comparison and evaluation choices weaken the evidence. The task formulation and method are plausible, but the manuscript needs additional experiments and clearer reporting before the core claim is fully convincing.

major comments (4)
  1. [§5.1–§5.2, Eq. (5)] The method’s core interaction-control mechanism is a text-to-mask-to-image bottleneck: T_inter is not an input to U_hoi in Eq. (5); the final image is conditioned only on F_detail (which includes M_fore) and F_ID. The paper does not test whether the binary union mask alone disambiguates interactions with similar silhouettes—e.g., 'feed' vs 'pet' vs 'hold'. Tab. 4 measures only KL divergence between mask distributions and ground truth, not per-prompt semantic correctness. Please add: (a) a comparison of MGIG with ground-truth masks vs IAMG-generated masks on final interaction accuracy; (b) an evaluation on confusable verb pairs to quantify semantic loss through the bottleneck. Without this, the central interaction-control claim is unsupported.
  2. [§6.3, Tab. 3, Tab. 4] The spatial-configuration alignment metric in Tab. 4 is a distributional KL divergence, not a per-sample semantic measure. A low KL can be achieved even if individual masks are wrong for the specific prompt. Moreover, the spatial mAP gap between InteractDiffusion (29.47) and our method (22.07) is dismissed as unfair because InteractDiffusion uses ground-truth boxes. This comparison is still presented in the main table and the gap is large. Please provide a fair comparison—e.g., InteractDiffusion with predicted boxes, or our method with ground-truth boxes/masks—or remove the direct comparison from the headline results. As presented, the 'spatial configuration' advantage over interaction-control baselines is not established.
  3. [§6.1, Tabs. 2–6] All quantitative results are reported as single numbers with no error bars, confidence intervals, or significance tests. Several differences are small—for example, the VLM accuracy difference between 'NIA + IA (Image)' (85.73) and 'NIA + IA (Video)' (85.75) is 0.02 percentage points, and the CLIP/DINO gains in Tab. 6 are also modest. These could be within run-to-run noise. Please report means and standard deviations over at least three seeds for the key tables, and indicate which differences are statistically significant (e.g., via paired bootstrap or t-test).
  4. [§6.4, Tab. 5] The user study is described as having 5 participants and 8 randomly selected samples, yielding 0.8K responses. This protocol is under-specified (how exactly 0.8K responses arise from 5×8?), and no inter-annotator agreement or statistical test is reported. Given the small sample, the statement that the results 'suggest its reliability' is overstated. Please either provide a more substantial user study with agreement metrics and significance testing, or temper the conclusion to reflect the limited evidence.
minor comments (4)
  1. [References / Tabs. 2–3] Citation numbering for MIP-Adapter is inconsistent: it appears as [13] in §1 and as [26] in Fig. 1 and Tabs. 2–3. Reference [26] actually corresponds to Subject-Diffusion. This needs correction throughout.
  2. [§5.2] Typos: 'tagret human object' should be 'target human object'; 'tha pattern' should be 'the pattern'. Also in §1, 'Inteact-Custom' should be 'Interact-Custom'.
  3. [Tab. 6] The column 'VLM judgement (%)' is not defined in the table or the surrounding text. Specify that this is the LLaVA holistic semantic accuracy introduced in §6.1.
  4. [§6.4] The user study's generation of 0.8K responses from 5 participants and 8 samples is not explained. Clarify the counting (e.g., 5 participants × 8 samples × 4 metrics × 5 models? = 800) and add error bars or per-metric distributions for the mean ratings.

Circularity Check

0 steps flagged

No significant circularity: the central claim is an empirical comparison validated by external detectors and VLMs; the only self-citations are related-work references that are not load-bearing.

full rationale

The paper's derivation chain is a two-stage pipeline: IAMG generates a foreground mask from the interaction text prompt (Eq. 1-4), and MGIG generates the image conditioned on that mask plus identity features (Eq. 5-6). Neither stage defines its output in terms of the final evaluation target, and the mask is an intermediate representation rather than a fitted parameter renamed as a prediction. Identity fidelity is measured with CLIP/DINO scores against segmented ground-truth regions; interaction semantics are measured with external detectors (RLIPv2, FGAHOI), an external VLM (LLaVA), and human ratings. These are not fitted by the paper's training procedure, so the headline numbers are not forced by construction. The self-citations (notably SA-HOI [38], and other works by the same group) appear only in related work and are not used to justify the method's architecture, to supply a uniqueness theorem, or to forbid alternative designs, so they are not load-bearing. The paper's own stated limitations—mask/subject shape conflicts handled by a two-stage training trick, a residual gap to real images, and lack of granular mask evaluation—are honest caveats about correctness and evaluation strength, not evidence of circularity. The use of Intern-VL to annotate prompts and LLaVA to judge semantic alignment involves different VLMs and measures the task's actual input-output relationship, so it does not reduce to a self-consistency check. Overall, no circular step can be exhibited from the paper's equations or citations; the claims are empirical and externally benchmarked.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

No new physical or conceptual entities are postulated; the 'foreground mask' is a representation, not an invented entity. There are no hand-fitted numeric constants in the central claim; the method is a trained neural pipeline whose learned weights are not free parameters. The main conceptual load is carried by the axioms listed above, especially the sufficiency of binary mask guidance and the quality of the automatically constructed dataset.

axioms (5)
  • domain assumption Existing HOI datasets (HICO-Det, V-COCO, Action Genome, VidHOI) are insufficient for CHOI because each human-object pair appears in only a static pose.
    Used to justify building a new dataset; asserted in Section 1 and Section 4 but not empirically demonstrated.
  • ad hoc to paper A binary foreground mask generated from the text prompt provides sufficient spatial configuration to express the interaction semantics.
    Core design choice in Sections 5.1 and 5.2; not proven, and the paper admits a quality gap between generated and real images.
  • domain assumption DINOv2 coarse features plus Sobel high-frequency maps are sufficient to preserve both human and object identity while allowing pose variation.
    Inherited from AnyDoor and applied to two subjects in Section 5.2 without independent validation for the new setting.
  • domain assumption Training samples with the same human-object pair in different poses teach the model to disentangle identity features from pose and interaction features.
    Central premise of the dataset design in Section 4; no analysis is shown that disentanglement actually occurs.
  • domain assumption Automated annotation tools (Grounded-SAM for masks, Intern-VL for action verbs) produce accurate enough labels for 1M training samples.
    Used in Section 4.2; no accuracy analysis or manual verification is reported.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Interact-Custom: Customized Human Object Interaction Image Generation." pith.science (2026). https://pith.science/paper/GS6CWGOB

@misc{pith2026250819575,
  author       = {Pith},
  title        = {Pith review of: Interact-Custom: Customized Human Object Interaction Image Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GS6CWGOB}},
  note         = {Machine review of arXiv:2508.19575}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Compositional Customized Image Generation aims to customize multiple target concepts within generation content, which has gained attention for its wild application. Existing approaches mainly concentrate on the target entity's appearance preservation, while neglecting the fine-grained interaction control among target entities. To enable the model of such interaction control capability, we focus on human object interaction scenario and propose the task of Customized Human Object Interaction Image Generation(CHOI), which simultaneously requires identity preservation for target human object and the interaction semantic control between them. Two primary challenges exist for CHOI:(1)simultaneous identity preservation and interaction control demands require the model to decompose the human object into self-contained identity features and pose-oriented interaction features, while the current HOI image datasets fail to provide ideal samples for such feature-decomposed learning.(2)inappropriate spatial configuration between human and object may lead to the lack of desired interaction semantics. To tackle it, we first process a large-scale dataset, where each sample encompasses the same pair of human object involving different interactive poses. Then we design a two-stage model Interact-Custom, which firstly explicitly models the spatial configuration by generating a foreground mask depicting the interaction behavior, then under the guidance of this mask, we generate the target human object interacting while preserving their identities features. Furthermore, if the background image and the union location of where the target human object should appear are provided by users, Interact-Custom also provides the optional functionality to specify them, offering high content controllability. Extensive experiments on our tailored metrics for CHOI task demonstrate the effectiveness of our approach.

Figures

Figures reproduced from arXiv: 2508.19575 by Yang Liu, Yuxin Peng, Zhaowen Wang, Zhu Xu.

Figure 1
Figure 1. Figure 1: Visualizations of interaction-involved generation [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Processing procedure for image and video data. For [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overall pipeline of Interact-Custom. In Interaction-Aware Mask Generation (IAMG) Stage, the Generation-Based [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of different approaches. Our approach yields better identity preservation ability as well as [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. MIBE: Multi-subject Interaction Benchmark and Evaluator for Personalized Image Generation

    cs.CV 2026-07 unverdicted novelty 6.0

    MIBE introduces a multi-subject interaction benchmark (MIB) with silver and gold sets and a dual-head evaluator (MIE) trained on VLM labels that outperforms baselines in matching human judgments.

Reference graph

Works this paper leans on

44 extracted references · 27 canonical work pages · cited by 1 Pith paper · 4 internal anchors

  1. [1]

    Yu-Wei Chao, Zhan Wang, Yugeng He, Jiaxuan Wang, and Jia Deng. 2015. HICO: A Benchmark for Recognizing Human-Object Interactions in Images. InProceedings of the IEEE International Conference on Computer Vision

    work page 2015

  2. [2]

    Hong Chen, Yipeng Zhang, Xin Wang, Xuguang Duan, Yuwei Zhou, and Wenwu Zhu. 2023. DisenBooth: Disentangled Parameter-Efficient Tuning for Subject- Driven Text-to-Image Generation. arXiv:2305.03374 (2023)

    Pith/arXiv arXiv 2023

  3. [3]

    Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Heng- shuang Zhao. 2024. AnyDoor: Zero-shot Object-level Image Customization. arXiv:2307.09481 [cs.CV] https://arxiv.org/abs/2307.09481

    Pith/arXiv arXiv 2024

  4. [4]

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. 2023. InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks. arXiv preprint arXiv:2312.14238 (2023)

    Pith/arXiv arXiv 2023

  5. [5]

    Meng-Jiun Chiou, Chun-Yu Liao, Li-Wei Wang, Roger Zimmermann, and Jiashi Feng. 2021. ST-HOI: A Spatial-Temporal Baseline for Human-Object Interaction Detection in Videos. In Proceedings of the 2021 Workshop on Intelligent Cross-Data Analysis and Retrieval. 9–17

    work page 2021

  6. [6]

    Frédéric Cordier, WonSook Lee, HyeWon Seo, and Nadia Magnenat-Thalmann

    work page

  7. [7]

    Ganggui Ding, Canyu Zhao, Wen Wang, Zhen Yang, Zide Liu, Hao Chen, and Chunhua Shen. 2024. FreeCustom: Tuning-Free Customized Image Generation for Multi-Concept Composition. arXiv:2405.13870 [cs.CV] https://arxiv.org/abs/ 2405.13870

    Pith/arXiv arXiv 2024

  8. [8]

    Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. 2023. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR

    work page 2023

  9. [9]

    Jiayi Gao, Zijin Yin, Changcheng Hua, Yuxin Peng, Kongming Liang, Zhanyu Ma, Jun Guo, and Yang Liu. 2025. ConMo: Controllable Motion Disentanglement and Recomposition for Zero-Shot Motion Transfer. arXiv preprint arXiv:2504.02451 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  10. [10]

    Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, et al . 2023. Mix-of- Show: Decentralized Low-Rank Adaptation for Multi-Concept Customization of Diffusion Models. In NeurIPS

    work page 2023

  11. [11]

    Saurabh Gupta and Jitendra Malik. 2015. Visual Semantic Role Labeling. arXiv preprint arXiv:1505.04474 (2015)

    Pith/arXiv arXiv 2015

  12. [12]

    Jiun Tian Hoe, Xudong Jiang, Chee Seng Chan, Yap-Peng Tan, and Weipeng Hu. 2024. InteractDiffusion: Interaction Control in Text-to-Image Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 6180–6189

    work page 2024

  13. [13]

    Qihan Huang, Siming Fu, Jinlong Liu, Hao Jiang, Yipeng Yu, and Jie Song. 2024. Resolving Multi-Condition Confusion for Finetuning-Free Personalized Image Generation. arXiv:2409.17920 [cs.CV] https://arxiv.org/abs/2409.17920

    Pith/arXiv arXiv 2024

  14. [14]

    Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. 2019. Ac- tion Genome: Actions as Composition of Spatio-temporal Scene Graphs. arXiv:1912.06992 [cs.CV] https://arxiv.org/abs/1912.06992

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  15. [15]

    Nick Kanopoulos, Nagesh Vasanthavada, and Robert L Baker. 1988. Design of an image edge detection filter using the Sobel operator. JSSC (1988)

    work page 1988

  16. [16]

    Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic opti- mization. arXiv:1412.6980 (2014)

    Pith/arXiv arXiv 2014

  17. [17]

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al

  18. [18]

    Ting Lei, Shaofeng Yin, and Yang Liu. 2024. Exploring the Potential of Large Foundation Models for Open-Vocabulary HOI Detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . 16657– 16667

    work page 2024

  19. [19]

    Ting Lei, Shaofeng Yin, Yuxin Peng, and Yang Liu. 2024. Exploring conditional multi-modal prompts for zero-shot hoi detection. In European Conference on Computer Vision. Springer, 1–19

    work page 2024

  20. [20]

    Yanshan Li, Tianyu Guo, Xing Liu, Wenhan Luo, and Weixin Xie. 2022. Action status based novel relative feature representations for interaction recognition. Chinese Journal of Electronics 31, 1 (2022), 168–180. doi:10.1049/cje.2020.00.088

    work page doi:10.1049/cje.2020.00.088 2022

  21. [21]

    Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. 2023. GLIGEN: Open-Set Grounded Text- to-Image Generation. CVPR (2023)

    work page 2023

  22. [22]

    Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, Ming-Ming Cheng, and Ying Shan. 2023. PhotoMaker: Customizing Realistic Human Photos via Stacked ID Embedding. arXiv:2312.04461 [cs.CV] https://arxiv.org/abs/2312.04461

    Pith/arXiv arXiv 2023

  23. [23]

    Zeyi Li, Pan Wang, and Zixuan Wang. 2024. FlowGANAnomaly: Flow-Based Anomaly Network Intrusion Detection with Adversarial Learning. Chinese Journal of Electronics 33, 1 (2024), 58–71. doi:10.23919/cje.2022.00.173

    work page doi:10.23919/cje.2022.00.173 2024

  24. [24]

    Zhuoying Li, Zhu Xu, Yuxin Peng, and Yang Liu. [n. d.]. Balancing Preservation and Modification: A Region and Semantic Aware Metric for Instruction-Based Image Editing. In Forty-second International Conference on Machine Learning

    work page

  25. [25]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual Instruc- tion Tuning. arXiv:2304.08485 [cs.CV] https://arxiv.org/abs/2304.08485

    Pith/arXiv arXiv 2023

  26. [26]

    Jian Ma, Junhao Liang, Chen Chen, and Haonan Lu. 2024. Subject-Diffusion:Open Domain Personalized Text-to-Image Generation without Test-time Fine-tuning. arXiv:2307.11410 [cs.CV] https://arxiv.org/abs/2307.11410

    Pith/arXiv arXiv 2024

  27. [27]

    Shuailei Ma, Yuefeng Wang, Shanze Wang, and Ying Wei. 2023. FGAHOI: Fine- Grained Anchors forHuman-Object Interaction Detection

    work page 2023

  28. [28]

    Qingchao Chen Yuxin Peng Minghang Zheng, Xinhao Cai and Yang Liu. 2024. Training Free Video Temporal Grounding using Large-scale Pre-trained Models. In Proceedings of the European Conference on Computer Vision (ECCV)

    work page 2024

  29. [29]

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El- Nouby, et al. 2024. Dinov2: Learning robust visual features without supervision. TMLR (2024)

    work page 2024

  30. [30]

    Yiming Qin, Zhu Xu, and Yang Liu. 2025. Apply Hierarchical-Chain-of-Generation to Complex Attributes Text-to-3D Generation. https://api.semanticscholar.org/ CorpusID:278481349

    work page 2025

  31. [31]

    Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, Zhaoyang Zeng, Hao Zhang, Feng Li, Jie Yang, Hongyang Li, Qing Jiang, and Lei Zhang. 2024. Grounded SAM: Assembling Open-World Models for Diverse Visual Tasks. arXiv:2401.14159 [cs.CV]

    Pith/arXiv arXiv 2024

  32. [32]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In CVPR

    2022

  33. [33]

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. 2023. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR

    work page 2023

  34. [34]

    Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, He Zhang, Wei Xiong, and Daniel Aliaga. 2024. IMPRINT: Gen- erative Object Compositing by Learning Identity-Preserving Representation. arXiv:2403.10701 [cs.CV] https://arxiv.org/abs/2403.10701

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  35. [35]

    Yoad Tewel, Rinon Gal, Gal Chechik, and Yuval Atzmon. 2024. Key-Locked Rank One Editing for Text-to-Image Personalization. arXiv:2305.01644 [cs.CV] https://arxiv.org/abs/2305.01644

    Pith/arXiv arXiv 2024

  36. [36]

    Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, Anthony Chen, Huaxia Li, Xu Tang, and Yao Hu. 2024. InstantID: Zero-shot Identity-Preserving Generation in Seconds. arXiv:2401.07519 [cs.CV] https://arxiv.org/abs/2401.07519

    Pith/arXiv arXiv 2024

  37. [37]

    Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. 2023. FastComposer: Tuning-Free Multi-Subject Image Generation with Localized Attention. arXiv:2305.10431 (2023)

    Pith/arXiv arXiv 2023

  38. [38]

    Zhu Xu, Qingchao Chen, Yuxin Peng, and Yang Liu. 2024. Semantic-Aware Human Object Interaction Image Generation. In Forty-first International Conference on Machine Learning

    work page 2024

  39. [39]

    Hangjie Yuan, Shiwei Zhang, Xiang Wang, Samuel Albanie, Yining Pan, Tao Feng, Jianwen Jiang, Dong Ni, Yingya Zhang, and Deli Zhao. 2023. RLIPv2: Fast Scaling of Relational Language-Image Pre-training. InProceedings of the IEEE/CVF International Conference on Computer Vision

    work page 2023

  40. [40]

    Bo Zhang, Yuxuan Duan, Jun Lan, Yan Hong, Huijia Zhu, Weiqiang Wang, and Li Niu. 2023. ControlCom: Controllable Image Composition using Diffusion Model. arXiv preprint arXiv:2308.10040 (2023)

    Pith/arXiv arXiv 2023

  41. [41]

    YE Zhaoda, HE Xiangteng, and PENG Yuxin. 2022. Unsupervised Cross-Media Hashing Learning via Knowledge Graph. Chinese Journal of Electronics 31, 6 (2022), 1081–1091. doi:10.1049/cje.2021.00.455

    work page doi:10.1049/cje.2021.00.455 2022

  42. [42]

    Chenyang Zhu, Kai Li, Yue Ma, Chunming He, and Li Xiu. 2024. Multi- Booth: Towards Generating All Your Concepts in an Image from Text. arXiv:2404.14239 [cs.CV] https://arxiv.org/abs/2404.14239

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  43. [2001]

    In People and Computers XV—Interaction without Frontiers: Joint Proceedings of HCI 2001 and IHM 2001

    From 2D photos of yourself to virtual try-on dress on the web. In People and Computers XV—Interaction without Frontiers: Joint Proceedings of HCI 2001 and IHM 2001. Springer, 31–46

    work page 2001

  44. [2023]

    Segment anything. In ICCV

    work page

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.