Pith. sign in

REVIEW 3 major objections 4 minor 48 references

LaRender: Training-Free Occlusion Control in Image Generation via Latent Rendering

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Volume-render the latent space to control occlusion in AI images

desk verdict Novel, training-free occlusion compositing that mostly works in the examples, but the evaluation under-supports the strong claims; worth a serious referee. read the letter →

arxiv 2508.07647 v1 pith:SWRDB44R submitted 2025-08-11 cs.CV

Xiaohang Zhan , Dingming Liu This is my paper
classification cs.CV
keywords imagegenerationdiffusionmodelocclusioncontrolvolumerenderinglatentspacetraining-freesemanticdensityattentionmap
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

The paper claims that a pre-trained text-to-image diffusion model can control which objects occlude which, with no training or fine-tuning, by replacing its cross-attention layers with a latent-space volume rendering formula. Each object's hidden features are treated as semi-transparent layers and composited front-to-back using transmittance estimated from bounding boxes and cross-attention maps. If correct, this lets users specify an occlusion graph and get images that respect the ordering, while also fading or strengthening objects by setting a per-object semantic opacity. The claim is supported by experiments on two datasets, where LaRender reports higher occlusion accuracy than text-to-image and layout-to-image baselines.

What carries the argument

The Latent Rendering formula (Eq. 3), a volume-rendering-style compositing operation applied to hidden states instead of RGB colors, together with per-object transmittance maps $M_i$ and a fast-to-slow density schedule $\sigma_i(t) = D_i T / (T+1-t)$. It replaces cross-attention and carries the entire occlusion-adjusting mechanism without learnable parameters.

What would settle it

Generate a set of prompts with two objects whose bounding boxes overlap strongly, render the same scene twice with the occlusion order swapped, and measure whether the output identity and ordering actually flip. If many swapped-order pairs produce nearly identical images, or if near-opaque settings ($\alpha \to 0.99$) still let background bleed through, the volume-rendering compositing is not physically encoding occlusion as claimed.

Watch

Extended reading notes

Core claim

LaRender modifies a latent diffusion model by swapping vanilla cross-attention for a Latent Rendering layer. For each layer, the latent attends separately to each object's prompt, producing object-wise features $R_i^{(l)}$; a step-dependent semantic density $\sigma_i$ controls how much each object participates, and accumulated transmittance $T_i = \exp(-\sum_{j<i} M_j \sigma_j)$ attenuates occluded objects. The rendered latent is a normalized, transmittance-weighted sum: $R^{(l+1)} = (1/S) \sum_i T_i(1-\exp(-\sigma_i)) M_i R_i^{(l)}$. The paper argues this physically ensures occlusion ordering because features behind an occluder are damped before they reach the virtual camera. The same opera

Load-bearing premise

The method assumes that object-wise latent features can be mixed with scalar semantic densities and a normalization factor to produce a valid next-layer latent without destroying object concepts; this is justified by qualitative results and a timing observation, not by derivation or a systematic failure analysis.

Editorial extensions

If this is right

  • If the central claim holds, occlusion control works in complex and unusual scenes with multiple occluding objects, using only a single hyperparameter $\alpha = 0.8$ in quantitative tests.
  • The mechanism is architecture-agnostic: results are shown on both SDXL and FLUX base models, suggesting it can be dropped into other latent diffusion backbones.
  • Adjusting semantic opacity extends beyond occlusion to visual effects: mass density changes (forests), particle concentration (fog, rain), light intensity, and lens-effect strength.
  • Inference cost stays close to the base model because the added operations are element-wise multiplications and sums that can be parallelized.
  • CLIP score degrades only slightly compared with SDXL (30.98 vs. 31.12), indicating the rendering layer does not strongly disturb the base model's generative quality.

Reading between the lines

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

  • The same latent-rendering layer could plausibly be inserted into video diffusion models or cascaded across frames to control temporal occlusion order, a natural extension the paper does not explore.
  • The attention-map refinement of transmittance suggests a possible fully box-free variant in which attention maps alone determine ordering, which could be tested by removing bounding-box inputs entirely.
  • The observed failure mode of lost or mixed concepts suggests a robustness boundary: a systematic sweep over overlap ratios and object counts could define where the scalar-density compositing assumption starts to break.
  • Because semantic density acts on arbitrary concepts (including non-objects like reflection or sunlight), the method may serve as a general-purpose concept-strength dial in generative pipelines.
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 / 4 minor

Summary. The paper proposes LaRender, a training-free method for controlling occlusion in text-to-image diffusion models. Given an occlusion graph and bounding boxes, LaRender topologically sorts objects, computes object-wise latent features at each cross-attention layer, derives transmittance maps from cross-attention maps and masks, and composites the latent features with a volume-rendering-inspired formula (Eq. 3) using user-set semantic densities. A step-dependent density schedule keeps features opaque early and increasingly mixed later. The method is evaluated on T2I-CompBench++ and a new RealOcc dataset against SDXL, FLUX, MIGC, and 3DIS, with UniDet and user-study metrics, and it demonstrates controllable opacity/density effects.

Significance. LaRender is a simple and elegant intervention: it introduces no learnable parameters, has negligible extra inference cost, works with pre-trained SDXL/FLUX, and goes beyond prompt-based occlusion control. The new RealOcc benchmark and the semantic-opacity applications are useful contributions. However, the central claim that the mechanism 'physically ensures' occlusion is not yet supported by the evidence: the automated metric is a single point estimate, the real-layout dataset is small, the user study is based on 15 selected samples, and the paper's own limitation section admits unquantified concept loss/mixing. With added statistical analysis and failure-rate reporting, this could be a solid contribution.

major comments (3)
  1. [§3.2, Eq. (3); §5, Fig. 6] The load-bearing assumption is that object-wise latent features R_i^(l) can be mixed as a normalized linear combination and still be valid inputs to the next layer. Eq. (3) is alpha compositing applied to latents; it is not a physical rendering of an actual scene. The paper justifies the density schedule only by a citation to [1] and shows qualitative results. The admitted failure mode in §5 — 'concepts can be lost or mixed' (Fig. 6) — is central, and its frequency is unreported. Without per-prompt success/failure statistics and basic latent-distribution diagnostics (e.g., drift of R^(l+1) relative to vanilla SDXL latents), the claim that occlusion relationships are 'physically ensured' is an overstatement. Please quantify and analyze failure rates.
  2. [§4.2, Table 1] The UniDet scores in Table 1 are point estimates with no error bars, per-prompt variance, or significance tests. The T2I-CompBench++ gap over the best baseline is 0.416 vs 0.401 (FLUX), which is within plausible noise. The user-study CIs are helpful but do not substitute for variance reporting on the automated metric. I request bootstrap or per-prompt confidence intervals, and ideally multiple sampling seeds, before the 'significantly outperforms' conclusion can be accepted. The CLIP-score comparison should also include variance.
  3. [§4.2, RealOcc, Table 1] RealOcc has only 70 images, and the user study uses 15 samples per dataset. Table 1 reports no UniDet or other automated occlusion metric for RealOcc, so the real-layout claim rests almost entirely on a 15-question user study with wide CIs (e.g., HPSR 0.767 ± 0.094). The paper should clarify how the 15 samples were selected, report per-sample agreement, and provide at least one automated occlusion metric on RealOcc using the COCOA amodal annotations. Without this, the RealOcc evaluation is too thin to support the cross-dataset claim.
minor comments (4)
  1. [§2.2] The sentence 'We are the first to address this challenge in the generative domain' is stronger than needed; the paper itself compares against layout-to-image methods that partially address occlusion. Suggest softening.
  2. [§3.2, Eq. (3)] The normalization S is undefined where all M_i are zero. State in the main text how backgrounds and empty regions are handled (e.g., a blank-prompt background object).
  3. [Table 2] The notation 'σ_i(t) = D_i T' for fixed opaque mode is ambiguous; clarify that this means σ_i(t) = D_i T for all t, whereas Eq. (4) gives σ_i(T) = D_i T.
  4. [§4.2] The reported LLM parsing accuracy (>95%) is given without a protocol or dataset. Please provide the evaluation set and prompt template, or mark this as an informal observation.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: LaRender's formula is a direct, explicitly acknowledged adaptation of volume rendering, with occlusion order and transmittance as inputs rather than fitted predictions.

full rationale

LaRender's central operation (Eq. 3) is an explicit adaptation of the NeRF quadrature formula (Eq. 2), with color c_i replaced by object-wise latent R_i^(l), density σ_i, and an additional mask M_i. The paper states 'we draw an analogy to Volume Rendering' and 'we formally define Latent Rendering'; it does not claim a new hidden derivation. The occlusion order and transmittance maps are user/LLM-supplied inputs; the formula deterministically composites front-to-back via T_i. Thus the output's occlusion order is by construction the input order, which is a control mechanism rather than a fitted prediction. The only mildly self-referential component is the transmittance map: M_i is computed from the same model's cross-attention maps ('we reuse the cross-attention maps from the aforementioned cross-attention module'). This is an inference-time design choice, not a training loop or a reduction of the output to a fitted parameter; the paper still checks end-to-end correctness with an external metric (UniDet) and user studies. The sole self-citation [40] supports the background claim 'depth does not equate to occlusion' and is not load-bearing for the method. The Limitations section's admission that 'concepts can be lost or mixed' is a robustness concern, not circularity. Therefore no circular step is established; score 2 reflects only the minor self-reference/self-citation.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The method's load-bearing choices are the per-object density D_i and three modeling assumptions: latent features are linearly composable, attention maps approximate transmittance, and the remaining pretrained network accepts composited latents. No new physical entities are introduced; the 'physics' is an analogy to alpha compositing.

free parameters (2)
  • semantic opacity alpha_i (semantic density D_i) = 0.8 for all quantitative comparisons; varied 0.1-0.9 in qualitative experiments
    User-chosen per-object scalar controlling the mixing weight in Eq. 3-4. Not fitted to data, but the method's success depends on choosing a reasonable value; Table 2 shows the density schedule affects performance.
  • Inverse-proportional density schedule constants = T follows base model default; denominator T+1-t
    The specific schedule in Eq. 4 is hand-designed and validated only by ablation; it is a modeling choice, not derived from first principles.
assumptions (4)
  • domain assumption Volume rendering quadrature (Eq. 2) is an appropriate model for compositing discrete object latents.
    Sections 3.1-3.2 assert that occlusion in images shares the essence of volume rendering; the paper does not prove that latent features satisfy the same physical laws. The central compositing rule rests on this analogy.
  • domain assumption The normalized weighted sum in Eq. 3 preserves the semantic identity of each object in latent space.
    Replacing cross-attention with a linear combination of object-wise latents assumes concepts survive mixing. The paper supports this only with qualitative results and a citation to concept-formation timing [1], not with a derivation.
  • domain assumption Cross-attention maps of the subject token, rescaled and masked by bounding boxes, give a usable estimate of transmittance.
    Section 3.2, Transmittance map paragraph. Attention maps encode where the model looks, not necessarily object occupancy; this approximation is load-bearing for where objects are composited.
  • domain assumption Stable Diffusion XL's remaining UNet layers can generate coherent images from recombination of object-wise latents after cross-attention is replaced.
    The method replaces every cross-attention layer, assuming the rest of the network accepts composited latents. Evidence is empirical only.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LaRender: Training-Free Occlusion Control in Image Generation via Latent Rendering." pith.science (2026). https://pith.science/paper/SWRDB44R

@misc{pith2026250807647,
  author       = {Pith},
  title        = {Pith review of: LaRender: Training-Free Occlusion Control in Image Generation via Latent Rendering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SWRDB44R}},
  note         = {Machine review of arXiv:2508.07647}
}
read the original abstract

We propose a novel training-free image generation algorithm that precisely controls the occlusion relationships between objects in an image. Existing image generation methods typically rely on prompts to influence occlusion, which often lack precision. While layout-to-image methods provide control over object locations, they fail to address occlusion relationships explicitly. Given a pre-trained image diffusion model, our method leverages volume rendering principles to "render" the scene in latent space, guided by occlusion relationships and the estimated transmittance of objects. This approach does not require retraining or fine-tuning the image diffusion model, yet it enables accurate occlusion control due to its physics-grounded foundation. In extensive experiments, our method significantly outperforms existing approaches in terms of occlusion accuracy. Furthermore, we demonstrate that by adjusting the opacities of objects or concepts during rendering, our method can achieve a variety of effects, such as altering the transparency of objects, the density of mass (e.g., forests), the concentration of particles (e.g., rain, fog), the intensity of light, and the strength of lens effects, etc.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 25 canonical work pages

  1. [1]

    ediff-i: Text-to-image dif- fusion models with an ensemble of expert denoisers

    Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji- aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image dif- fusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 4

    arXiv 2022

  2. [2]

    Multidiffusion: Fusing diffusion paths for controlled image generation

    Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. 2023. 2

    2023

  3. [3]

    pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis

    Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799–5809, 2021. 3

    work page 2021

  4. [4]

    Efficient geometry-aware 3d generative adversarial networks

    Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16123–16133, 2022. 3

    work page 2022

  5. [5]

    Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis

    Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 2

    arXiv 2023

  6. [6]

    Training-free layout control with cross-attention guidance

    Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In Proceedings of the IEEE/CVF winter conference on applications of com- puter vision, pages 5343–5353, 2024. 2

    work page 2024

  7. [7]

    Layerfusion: Harmo- nized multi-layer text-to-image generation with generative priors

    Yusuf Dalva, Yijun Li, Qing Liu, Nanxuan Zhao, Jianming Zhang, Zhe Lin, and Pinar Yanardag. Layerfusion: Harmo- nized multi-layer text-to-image generation with generative priors. arXiv preprint arXiv:2412.04460, 2024. 3

    arXiv 2024

  8. [8]

    V ol- ume rendering

    Robert A Drebin, Loren Carpenter, and Pat Hanrahan. V ol- ume rendering. ACM Siggraph Computer Graphics , 22(4): 65–74, 1988. 3

    work page 1988

Show all 48 references
  1. [9]

    Stylenerf: A style-based 3d-aware genera- tor for high-resolution image synthesis

    Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware genera- tor for high-resolution image synthesis. arXiv preprint arXiv:2110.08985, 2021. 3

    2021 arXiv

  2. [10]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. 5

    2025 arXiv

  3. [11]

    Viewdiff: 3d-consistent image generation with text-to-image models

    Lukas H ¨ollein, Aljaˇz Boˇziˇc, Norman M¨uller, David Novotny, Hung-Yu Tseng, Christian Richardt, Michael Zollh ¨ofer, and Matthias Nießner. Viewdiff: 3d-consistent image generation with text-to-image models. In Proceedings of the IEEE/CVF conference on computer vision and pa...

    2024

  4. [12]

    spacy 2: Natural language understanding with bloom embeddings, convolutional neural networks and incremental parsing

    Matthew Honnibal. spacy 2: Natural language understanding with bloom embeddings, convolutional neural networks and incremental parsing. (No Title), 2017. 3

    2017

  5. [13]

    T2i-compbench++: An enhanced and comprehensive benchmark for compositional text-to-image generation

    Kaiyi Huang, Chengqi Duan, Kaiyue Sun, Enze Xie, Zhen- guo Li, and Xihui Liu. T2i-compbench++: An enhanced and comprehensive benchmark for compositional text-to-image generation. IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 2025. 5

    2025

  6. [14]

    Image gener- ation from scene graphs

    Justin Johnson, Agrim Gupta, and Li Fei-Fei. Image gener- ation from scene graphs. In Proceedings of the IEEE con- ference on computer vision and pattern recognition , pages 1219–1228, 2018. 2

    2018

  7. [15]

    Deep occlusion- aware instance segmentation with overlapping bilayers

    Lei Ke, Yu-Wing Tai, and Chi-Keung Tang. Deep occlusion- aware instance segmentation with overlapping bilayers. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 4019–4028, 2021. 3

    2021

  8. [16]

    Black Forest Labs. Flux. https://github.com/ black-forest-labs/flux, 2024. 2, 5, 6

    2024

  9. [17]

    Gligen: Open-set grounded text-to-image generation

    Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jian- wei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22511–22521, 2023. 2

    2023

  10. [18]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision–ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceeding...

    2014

  11. [19]

    Variational amodal object completion

    Huan Ling, David Acuna, Karsten Kreis, Seung Wook Kim, and Sanja Fidler. Variational amodal object completion. Advances in Neural Information Processing Systems , 33: 16246–16257, 2020. 3

    2020

  12. [20]

    Object-level scene deocclusion

    Zhengzhe Liu, Qing Liu, Chirui Chang, Jianming Zhang, Daniil Pakhomov, Haitian Zheng, Zhe Lin, Daniel Cohen-Or, and Chi-Wing Fu. Object-level scene deocclusion. In ACM SIGGRAPH 2024 Conference Papers, pages 1–11, 2024. 3

    2024

  13. [21]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. Communications of the ACM , 65(1):99–106, 2021. 3

    2021

  14. [22]

    Glide: Towards photorealistic image generation and editing with text-guided diffusion models.arXiv preprint arXiv:2112.10741, 2021

    Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models.arXiv preprint arXiv:2112.10741, 2021. 2

    2021 arXiv

  15. [23]

    Giraffe: Represent- ing scenes as compositional generative neural feature fields

    Michael Niemeyer and Andreas Geiger. Giraffe: Represent- ing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 11453–11464, 2021. 3

    2021

  16. [24]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF inter- national conference on computer vision , pages 4195–4205,

  17. [25]

    Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M ¨uller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion mod- els for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 5, 6

    2023 arXiv

  18. [26]

    Amodal instance segmentation with kins dataset

    Lu Qi, Li Jiang, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Amodal instance segmentation with kins dataset. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3014–3023, 2019. 3 9

    2019

  19. [27]

    Zero-shot text-to-image generation

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea V oss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International confer- ence on machine learning, pages 8821–8831. Pmlr, 2021. 2

    2021

  20. [28]

    Hierarchical text-conditional image gener- ation with clip latents

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gener- ation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022

    2022 arXiv

  21. [29]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 2

    2022

  22. [30]

    Photorealistic text-to-image diffusion models with deep language understanding

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information...

    2022

  23. [31]

    Graf: Generative radiance fields for 3d-aware im- age synthesis

    Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware im- age synthesis. Advances in Neural Information Processing Systems, 33:20154–20166, 2020. 3

    2020

  24. [32]

    Using scene graph context to improve im- age generation

    Subarna Tripathi, Anahita Bhiwandiwalla, Alexei Bastidas, and Hanlin Tang. Using scene graph context to improve im- age generation. arXiv preprint arXiv:1901.03762, 2019. 2

    1901 arXiv

  25. [33]

    Mulan: A multi layer anno- tated dataset for controllable text-to-image generation

    Petru-Daniel Tudosiu, Yongxin Yang, Shifeng Zhang, Fei Chen, Steven McDonagh, Gerasimos Lampouras, Ignacio Iacobacci, and Sarah Parisot. Mulan: A multi layer anno- tated dataset for controllable text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vi...

    2024

  26. [34]

    Spotactor: Training-free layout- controlled consistent image generation

    Jiahao Wang, Caixia Yan, Weizhan Zhang, Haonan Lin, Mengmeng Wang, Guang Dai, Tieliang Gong, Hao Sun, and Jingdong Wang. Spotactor: Training-free layout- controlled consistent image generation. arXiv preprint arXiv:2409.04801, 2024. 2

    2024 arXiv

  27. [35]

    Enhancing mmdit-based text-to-image models for similar subject generation

    Tianyi Wei, Dongdong Chen, Yifan Zhou, and Xingang Pan. Enhancing mmdit-based text-to-image models for similar subject generation. arXiv preprint arXiv:2411.18301, 2024. 8

    2024 arXiv

  28. [36]

    Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion

    Jinheng Xie, Yuexiang Li, Yawen Huang, Haozhe Liu, Wen- tian Zhang, Yefeng Zheng, and Mike Zheng Shou. Boxdiff: Text-to-image synthesis with training-free box-constrained diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7452–7461, 2023. 2

    2023

  29. [37]

    Diffusion-based scene graph to image gener- ation with masked contrastive pre-training

    Ling Yang, Zhilin Huang, Yang Song, Shenda Hong, Guohao Li, Wentao Zhang, Bin Cui, Bernard Ghanem, and Ming- Hsuan Yang. Diffusion-based scene graph to image gener- ation with masked contrastive pre-training. arXiv preprint arXiv:2211.11138, 2022. 2

    2022 arXiv

  30. [38]

    A tri- layer plugin to improve occluded detection

    Guanqi Zhan, Weidi Xie, and Andrew Zisserman. A tri- layer plugin to improve occluded detection. arXiv preprint arXiv:2210.10046, 2022. 3

    2022 arXiv

  31. [39]

    Amodal ground truth and completion in the wild

    Guanqi Zhan, Chuanxia Zheng, Weidi Xie, and Andrew Zis- serman. Amodal ground truth and completion in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28003–28013, 2024. 3

    2024

  32. [40]

    Self-supervised scene de- occlusion

    Xiaohang Zhan, Xingang Pan, Bo Dai, Ziwei Liu, Dahua Lin, and Chen Change Loy. Self-supervised scene de- occlusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3784–3792,

  33. [41]

    Transparent image layer diffusion using latent transparency

    Lvmin Zhang and Maneesh Agrawala. Transparent image layer diffusion using latent transparency. arXiv preprint arXiv:2402.17113, 2024. 3

    2024 arXiv

  34. [42]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3836–3847, 2023. 5

    2023

  35. [43]

    Iter- comp: Iterative composition-aware feedback learning from model gallery for text-to-image generation

    Xinchen Zhang, Ling Yang, Guohao Li, Yaqi Cai, Jiake Xie, Yong Tang, Yujiu Yang, Mengdi Wang, and Bin Cui. Iter- comp: Iterative composition-aware feedback learning from model gallery for text-to-image generation. arXiv preprint arXiv:2410.07171, 2024. 5

    2024 arXiv

  36. [44]

    Layoutdiffusion: Controllable diffu- sion model for layout-to-image generation

    Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Layoutdiffusion: Controllable diffu- sion model for layout-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 22490–22499, 2023. 2

    2023

  37. [45]

    Migc: Multi-instance generation controller for text-to-image synthesis

    Dewei Zhou, You Li, Fan Ma, Xiaoting Zhang, and Yi Yang. Migc: Multi-instance generation controller for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 6818– 6828, 2024. 2, 5, 6

    2024

  38. [46]

    3dis: Depth-driven decoupled instance synthesis for text-to-image generation

    Dewei Zhou, Ji Xie, Zongxin Yang, and Yi Yang. 3dis: Depth-driven decoupled instance synthesis for text-to-image generation. arXiv preprint arXiv:2410.12669, 2024. 2, 5, 6

    2024

  39. [47]

    Human de-occlusion: Invisible perception and recovery for humans

    Qiang Zhou, Shiyin Wang, Yitong Wang, Zilong Huang, and Xinggang Wang. Human de-occlusion: Invisible perception and recovery for humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 3691–3701, 2021. 3

    2021

  40. [48]

    #!#"$!$" $%&

    Yan Zhu, Yuandong Tian, Dimitris Metaxas, and Piotr Doll´ar. Semantic amodal segmentation. In Proceedings of the IEEE conference on computer vision and pattern recog- nition, pages 1464–1472, 2017. 3, 5 10 A. The RealOCC Dataset Statistics. As shown in Figure 7, the RealOCC da...

    2017

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.