Pith. sign in

REVIEW 3 major objections 7 minor 27 references

Diffusion models drop and merge objects in dense occluded scenes because attention leaks across instances and training underweights hidden ones; layout-ownership biases plus amodal-ratio loss reweighting fix both.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-12 02:08 UTC pith:4AJJYCJR

load-bearing objection Solid engineering fix for dense-layout counting: ownership-aware attention bias plus amodal loss reweighting, plus a useful dense benchmark; absolute numbers stay modest and transfer rests on synthetic proxies. the 3 major comments →

arxiv 2607.03488 v1 pith:4AJJYCJR submitted 2026-07-03 cs.CV

Learning to Generate Multiple Objects from Dense and Occluded Layouts

Bach-Hoang Ngo , Si-Tri Ngo , Hieu Le , Trung-Nghia Le This is my paper
classification cs.CV
keywords text-to-image diffusionobject countinglayout-to-imageocclusionamodal masksattention biasinstance ownershipOverlapDepth-45K
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.

Text-to-image diffusion models look fine on sparse prompts yet systematically fail when many objects overlap: nearby instances fuse into blobs and heavily occluded ones vanish, so the final count is wrong even when the layout boxes are correct. The authors locate the failure in two places: free self-attention lets features leak across overlapping regions, and the ordinary reconstruction loss gives almost no gradient to objects that occupy only a few visible pixels. They therefore inject soft, parameter-free attention biases that reward tokens belonging to the same depth-ordered box and repel tokens of different boxes, while reweighting each instance’s loss by the ratio of its full amodal area to its visible area (clipped and mean-normalized so the overall scale stays constant). A new synthetic benchmark of 45 000 dense multi-object scenes with amodal masks lets them measure the effect; the combined method cuts large counting errors by roughly a third relative to layout-only training and raises exact-count accuracy on both their own test set and public counting benchmarks, without needing masks at inference.

Core claim

Instance ownership collapse—cross-attention leakage among overlapping tokens plus weak gradients on occluded objects—is the dominant reason diffusion models produce the wrong object counts in dense layouts; soft layout-ownership attention biases together with an amodal-to-visible area reweighting of the diffusion loss restore separable, countable instances while preserving image quality.

What carries the argument

Amodal-Aware Instance-Balanced Loss (AIBL) plus overlap-ownership attention bias: each training instance is scaled by a clipped power of its amodal/visible area ratio, the resulting pixel weight map is mean-normalized, and the same depth-ordered region map supplies cohesion and separation additives inside the transformer’s attention logits.

Load-bearing premise

Depth-ordered box ownership and synthetic amodal-to-visible area ratios are faithful enough proxies that the same soft biases and loss weights will transfer to open-world photos and imperfectly inferred layouts.

What would settle it

On a held-out set of real photographs containing 15–25 heavily overlapping objects of known ground-truth counts, replace the synthetic amodal masks with human-annotated amodal masks (or none) and measure whether exact-match accuracy and RMSE remain better than the layout-only baseline by a comparable margin.

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

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

3 major / 7 minor

Summary. The paper argues that text-to-image diffusion models fail at dense multi-object counting because of instance ownership collapse: free cross-instance attention mixes overlapping tokens, while uniform reconstruction losses starve heavily occluded instances of gradient. It proposes two complementary fixes for DiT-style models: (i) a parameter-free layout attention bias that uses depth-ordered region identities to reinforce cohesion among ownership-winner tokens and separate overlapping instances (Eqs. 1–3), and (ii) Amodal-Aware Instance-Balanced Loss (AIBL), which reweights the flow-matching objective by amodal-to-visible area ratios with mean normalization (Eqs. 4–7), used only at training. A new synthetic benchmark, OverlapDepth-45K (45K dense scenes with amodal/depth metadata), is introduced. On that test set the full model reduces RMSE from 9.26 (layout-only) to 5.99 and improves exact/tolerance accuracy; it also outperforms CountGen and other baselines on CoCoCount and T2I-Compbench counting (including a 100-image human eval).

Significance. If the results hold under stronger evaluation, the work is a useful contribution to controllable dense compositional generation. The problem framing (ownership collapse as joint attention leakage + gradient starvation) is clear and actionable. Strengths include: a soft, non-trainable attention bias that avoids hard masking and training–inference mismatch; a training-only amodal reweighting scheme with explicit mean normalization so global gradient scale is preserved; clean component ablations (Table 3) plus γ-sensitivity and density-stratified analyses (Fig. 4); and external gains with human counting on Compbench. The OverlapDepth-45K resource is a concrete addition for studying severe overlap. Absolute exact-match remains modest, so impact is more incremental than transformative, but the design is practical and the empirical story is coherent.

major comments (3)
  1. [Section 4.2, Tables 1 and 3] The headline quantitative claim—35.3% RMSE reduction from 9.26 to 5.99 and the Exact/±10%/±20% gains in Tables 1 and 3—rests entirely on automated SAM-3 counting over the synthetic OverlapDepth-45K test set (Sec. 4.2). The text says detections were validated against human annotations on a subset, but no agreement statistics (e.g., per-category correlation, human–SAM MAE/RMSE, or confusion rates under heavy occlusion) are reported. Without those numbers, it is hard to know how much of the reported drop in catastrophic merging is true instance recovery versus detector bias on the same synthetic distribution the model was trained on. Please report quantitative human–SAM agreement on a stratified dense subset and, if agreement is imperfect, recompute primary metrics with human counts or a dual-detector protocol.
  2. [Sections 3.2, 4.1, 4.5; Eq. (4)] AIBL (Eq. 4) and the ownership map O (Eq. 5) assume clean amodal extents, visible areas, and depth-ordered ownership derived from synthetic masks (Sec. 3.2, 4.1). Real photos and LLM-inferred layouts lack this supervision; estimated amodal masks are noisy. There is no sensitivity study that corrupts α_i (noise on A_amodal / A_visible, wrong depth order, or missing instances) and remeasures counting. Sec. 4.5 acknowledges synthetic domain shift, but the central transfer claim—that the same recipe prevents merging under open-world occlusion—needs at least one such robustness check, or an experiment that trains AIBL with predicted rather than ground-truth amodal masks. Otherwise the 35% RMSE figure may overstate gains available outside OverlapDepth-45K.
  3. [Abstract; Table 3; Figure 4a; Section 4.5] Even after all components, Exact Match is only 15.32% and counting score collapses in the 23+ object bin (Fig. 4a: 8.3%, slightly below layout-only). The abstract and contributions state that the approach “substantially improves count accuracy and prevents instance merging.” That is true relative to the layout-only baseline on RMSE, but absolute fidelity remains low and extreme density is essentially unsolved (as the paper notes under density saturation). Please qualify the claim more carefully in the abstract/intro/conclusion, and discuss whether residual failures are primarily latent-resolution limits versus incomplete ownership modeling, with any supporting evidence (e.g., failure rate vs. box overlap fraction).
minor comments (7)
  1. [Section 1 contributions; Section 3.1–3.2] The contribution list calls layout attention “parameter-free,” which correctly means no new trainable weights, but ω_intra, ω_inter, β, the fixed 1/2 Δ scale, Gaussian σ, λ_bg, α_max, and γ are free hyperparameters. State the chosen values and any tuning protocol in the main text or appendix.
  2. [Figure 2; Eq. (1)] Figure 2 caption refers to an attention bias matrix M in one place and B in the overview text; Eq. (1) uses M. Unify notation (B vs M) across figure and equations.
  3. [Section 4.1] OverlapDepth-45K construction is underspecified: how amodal masks and ordinal depth are obtained or synthesized, rendering engine, and whether train/test share assets. A short construction paragraph or appendix would make the benchmark reusable.
  4. [Section 4.4] Ablation text says “identical LoRA rank” but never states the rank; learning rate and step count are given. Report rank and which DiT blocks receive LoRA.
  5. [Table 2; Section 4.3] Table 2 reports baseline numbers “as provided by Binyamin et al.” Clarify whether AIBL used the same backbone/protocol as CountGen or a different generator with layout conditioning, so the comparison is not confounded by layout vs. pure T2I.
  6. [Section 3.1, Eqs. (1)–(2)] In Eq. (1)–(2), the conditions for Δ_coh / Δ_sep are slightly dense; a one-line pseudocode or small diagram of contested-zone token pairs would help readers implement the bias.
  7. [Figure 1] Figure 1 bottom row reports 24/28 for the densest scene; the caption claims “exact or near-exact counts across all density levels.” Prefer consistent language (near-exact) where counts are not exact.

Circularity Check

0 steps flagged

No circularity: empirical method with independent counting metrics; no derivation reduces by construction to its inputs.

full rationale

The paper proposes two engineering components (layout-aware attention bias M from depth-ordered region identity R and ownership, Eqs. 1-2; AIBL reweighting via amodal-to-visible ratios α_i, Eq. 4, mean-normalized into W_norm) and validates them empirically on counting metrics (exact match, MAE/RMSE via SAM-3, human accuracy). These metrics are computed from generated images and external detectors/human counts; they do not algebraically equal the training weights, the painter's-algorithm ownership map, or any fitted hyperparameter (γ, β, ω). Ablations (Table 3) and external benchmarks (CoCoCount, T2I-Compbench) supply independent evidence. Hyperparameter selection on the authors' synthetic set is ordinary ML practice, not a self-definitional loop or fitted-input-as-prediction. No load-bearing self-citation, uniqueness theorem, or renamed known result appears. The derivation chain is therefore self-contained and non-circular.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 3 invented entities

The central empirical claim rests on standard diffusion/flow-matching practice, the modeling assumption that depth-ordered box ownership plus soft attention biases suffice to control feature leakage, and several hand-chosen scalars that shape the bias and loss. No new physical entities are postulated; OverlapDepth-45K is a constructed dataset rather than a fundamental entity. Free parameters are the usual method knobs whose values are selected for best reported metrics.

free parameters (6)
  • dampening exponent gamma = 0.5 (optimal among {0,0.3,0.5,0.75,1.0})
    Controls aggressiveness of occlusion boost in Eq. 4; swept and selected at 0.5 for best RMSE/MAE/exact match on the test set.
  • alpha_max clip
    Caps extreme occlusion boosts; value not numerically specified beyond existence of the clip.
  • omega_intra, omega_inter, beta, Delta scale 1/2 = beta>1; Delta multiplies by 1/2; exact omega values unspecified
    Base cohesion/separation biases and dense-scene grounding factor in Eqs. 1-2; stated as fixed, not tuned across experiments, but still design choices that define the method.
  • Gaussian soft-prior sigma
    Spatial decay width for M_soft in Eq. 3; not numerically reported.
  • lambda_bg background weight = 0.8
    Downweight for background pixels in the AIBL map (Eq. 5).
  • LoRA rank, lr=5e-5, 3000 steps = lr 5e-5, 3000 steps
    Training schedule shared across ablations; chosen by authors for the reported runs.
axioms (6)
  • domain assumption Rectified flow matching with velocity prediction is a valid training objective for the DiT backbone.
    Section 3.3 adopts Lipman et al. flow matching without re-derivation.
  • domain assumption Depth-ordered painter's algorithm on bounding boxes yields a correct ownership map for latent tokens in overlap zones.
    Section 3.1 region identity map; load-bearing for both attention bias and AIBL ownership.
  • ad hoc to paper Amodal-to-visible area ratio is a sufficient scalar proxy for the gradient starvation of occluded instances.
    Section 3.2 defines alpha_i from areas only; no proof that area ratio matches true optimization difficulty.
  • domain assumption Mean-normalizing the weight map preserves global gradient scale so learning rate needs no retuning.
    Eq. 6 and surrounding text; standard reweighting practice but assumed rather than measured across densities.
  • domain assumption Soft additive attention biases do not destroy the pretrained generative prior.
    Claimed in Section 3.1; supported only by qualitative quality preservation.
  • domain assumption SAM-3 zero-shot detections (with partial human check) are accurate enough to score counting fidelity.
    Section 4.2 evaluation protocol.
invented entities (3)
  • OverlapDepth-45K dataset no independent evidence
    purpose: Provide dense multi-object scenes with amodal masks and ordinal depth for training and evaluating count fidelity under occlusion.
    New benchmark constructed by the authors; independent evidence will exist only if the data is publicly released and reused by others.
  • Amodal-Aware Instance-Balanced Loss (AIBL) no independent evidence
    purpose: Reweight per-pixel diffusion loss by occlusion boost factors so occluded instances receive stronger gradients.
    Named training objective introduced in the paper; it is a loss design, not a physical entity, and is only evidenced by the paper's own ablations.
  • Overlap-ownership attention bias (Delta_coh / Delta_sep) no independent evidence
    purpose: Extra cohesion for ownership-winner tokens in contested zones and region-level separation to stop cross-instance leakage.
    Specific bias schedule defined in Eqs. 1-2; no external validation outside this work.

pith-pipeline@v1.1.0-grok45 · 16365 in / 3853 out tokens · 27287 ms · 2026-07-12T02:08:18.786128+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Learning to Generate Multiple Objects from Dense and Occluded Layouts." pith.science (2026). https://pith.science/paper/4AJJYCJR

@misc{pith2026260703488,
  author       = {Pith},
  title        = {Pith review of: Learning to Generate Multiple Objects from Dense and Occluded Layouts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4AJJYCJR}},
  note         = {Machine review of arXiv:2607.03488}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Text-to-image diffusion models fail to generate correct object counts in dense scenes, where overlapping instances collapse into indistinguishable structures despite appearing visually plausible. We identify this as instance ownership collapse: tokens from overlapping objects interact freely through attention, while heavily occluded instances receive weak supervision due to their small visible areas. We address this through layout-aware attention biases that softly bias token interactions toward region-consistent grouping and suppress cross-instance leakage, paired with an amodal-balanced loss that amplifies gradients for occluded objects based on their occlusion level. To enable systematic evaluation, we introduce OverlapDepth-45K, a benchmark of densely overlapping scenes with amodal supervision. Our approach substantially improves count accuracy and prevents instance merging while preserving image quality. Project page: https://bachngoh.github.io/AIBL

Figures

Figures reproduced from arXiv: 2607.03488 by Bach-Hoang Ngo, Hieu Le, Si-Tri Ngo, Trung-Nghia Le.

Figure 1
Figure 1. Figure 1: Count-preserving multi-category compositional generation. Standard diffusion models (top) succeed on simple prompts (3/3 swans) but fail as scene complexity grows - overcounting single-category scenes (7/6 apples), dropping instances in multi-category compositions (7/12 across bottles, plates, and glasses), and catastrophically merging objects in dense scenes (50+/28 across persons and cans). Our approach … view at source ↗
Figure 2
Figure 2. Figure 2: Architecture Overview. (a) Layout Attention (LA) is integrated into the transformer blocks of a DiT backbone to steer the generation process. (b) LA utilizes a region identity map to compute an attention bias matrix M, where cohesion and separation biases are applied to tokens based on their instance ownership. (c) The Amodal-Aware Instance-Balanced Loss (AIBL) module (training only) computes per-instance … view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison on counting prompts. Even with explicitly grounded bounding boxes, the layout-attention baseline (Base) frequently merges nearby instances. Our Amodal-Aware Instance-Balanced Loss (AIBL) rebalances training to promote clearer instance separation, producing more distinct objects (Ours). We also include reference generations from SDXL, FLUX.1-dev [15], and Z-Image [4] on the same promp… view at source ↗
Figure 4
Figure 4. Figure 4: Analysis of AIBL behavior. (Left) Counting score stratified by target object density. AIBL provides significant improvements in mid-to-high density regimes (13–22 objects). (Right) RMSE, MAE, and Exact Match accuracy as a function of the dampening exponent γ. Dotted lines indicate the layout-only baseline. The optimal γ=0.5 achieves the best trade-off across all metrics. larger values preserve the raw amod… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

27 extracted references · 5 linked inside Pith

  1. [1]

    Seethrough3d: Occlusion aware 3d control in text-to-image generation

    Vaibhav Agrawal, Rishubh Parihar, Pradhaan Bhat, Ravi Kiran Sarvadevabhatla, and R Venkatesh Babu. Seethrough3d: Occlusion aware 3d control in text-to-image generation. arXiv preprint arXiv:2602.23359, 2026

    arXiv 2026

  2. [2]

    Improving image generation with better captions

    James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn. openai. com/papers/dall-e-3. pdf, 2(3):8, 2023

    2023

  3. [3]

    Make it count: Text-to-image generation with an accurate number of objects

    Lital Binyamin, Yoad Tewel, Hilit Segev, Eran Hirsch, Royi Rassin, and Gal Chechik. Make it count: Text-to-image generation with an accurate number of objects. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

    2025

  4. [4]

    Z-image: An efficient image generation foundation model with single-stream diffusion transformer.arXiv preprint arXiv:2511.22699, 2025

    Huanqia Cai, Sihan Cao, Ruoyi Du, Peng Gao, Steven Hoi, Zhaohui Hou, Shijie Huang, Dengyang Jiang, Xin Jin, Liangchen Li, et al. Z-image: An efficient image generation foundation model with single-stream diffusion transformer.arXiv preprint arXiv:2511.22699, 2025

    Pith/arXiv arXiv 2025

  5. [5]

    Sam 3: Segment anything with concepts.arXiv preprint arXiv:2511.16719, 2025

    Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoubhik Debnath, Ronghang Hu, Didac Suris, Chaitanya Ryali, Kalyan Vasudev Alwala, Haitham Khedr, Andrew Huang, et al. Sam 3: Segment anything with concepts.arXiv preprint arXiv:2511.16719, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models.ACM Trans

    Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models.ACM Trans. Graph., 42(4), July 2023

    2023

  7. [7]

    Training-free layout control with cross-attention guidance

    Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024

    2024

  8. [8]

    Reason out your layout: Evoking the layout master from large language models for text-to-image synthesis.arXiv preprint arXiv:2311.17126, 2023

    Xiaohui Chen, Yongfei Liu, Yingxiang Yang, Jianbo Yuan, Quanzeng You, Li-Ping Liu, and Hongxia Yang. Reason out your layout: Evoking the layout master from large language models for text-to-image synthesis.arXiv preprint arXiv:2311.17126, 2023

    Pith/arXiv arXiv 2023

  9. [9]

    Prompt-to-prompt image editing with cross attention control.arXiv preprint arXiv:2208.01626, 2022

    Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control.arXiv preprint arXiv:2208.01626, 2022

    Pith/arXiv arXiv 2022

  10. [10]

    T2i-compbench: A compre- hensive benchmark for open-world compositional text-to-image generation.Advances in Neural Information Processing Systems, 36:78723–78747, 2023

    Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A compre- hensive benchmark for open-world compositional text-to-image generation.Advances in Neural Information Processing Systems, 36:78723–78747, 2023

    2023

  11. [11]

    Counting guidance for high fidelity text-to-image synthesis

    Wonjun Kang, Kevin Galim, Hyung Il Koo, and Nam Ik Cho. Counting guidance for high fidelity text-to-image synthesis. In2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 899–908. IEEE, 2025

    2025

  12. [12]

    Counting guidance for high fidelity text-to-image synthesis

    Wonjun Kang and Kevin Park. Counting guidance for high fidelity text-to-image synthesis. arXiv preprint arXiv:2306.17567

    Pith/arXiv arXiv

  13. [13]

    Omg: Occlusion-friendly personalized multi-concept generation in diffusion models

    Zhe Kong, Yong Zhang, Tianyu Yang, Tao Wang, Kaihao Zhang, Bizhu Wu, Guanying Chen, Wei Liu, and Wenhan Luo. Omg: Occlusion-friendly personalized multi-concept generation in diffusion models. InComputer Vision – ECCV 2024: 18th European Conference, Milan, Italy, September 29–October 4, 2024, Proceedings, Part XXXI, page 253–270, Berlin, Heidelberg,

    2024

  14. [14]

    Visual genome: Connecting language and vision using crowdsourced dense image annotations.International Journal of Computer Vision, 123(1):32–73, 2017

    Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations.International Journal of Computer Vision, 123(1):32–73, 2017

    2017

  15. [15]

    Flux.https://github.com/black-forest-labs/flux, 2024

    Black Forest Labs. Flux.https://github.com/black-forest-labs/flux, 2024. 10

    2024

  16. [16]

    Armand, Divyansh Srivastava, Xiaojun Shan, Zeyuan Chen, Jianwen Xie, and Zhuowen Tu

    Bingnan Li, Chen-Yu Wang, Haiyang Xu, Xiang Zhang, Ethan J. Armand, Divyansh Srivastava, Xiaojun Shan, Zeyuan Chen, Jianwen Xie, and Zhuowen Tu. Overlaybench: A benchmark for layout-to-image generation with dense overlaps. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2025

    2025

  17. [17]

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

    Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. GLIGEN: Open-set grounded text-to-image generation. InCVPR, 2023

    2023

  18. [18]

    Microsoft COCO: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. InEuropean Conference on Computer Vision. Springer, 2014

    2014

  19. [19]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. InICLR, 2023

    2023

  20. [20]

    T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models

    Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. InProceedings of the AAAI conference on artificial intelligence, volume 38, pages 4296–4304, 2024

    2024

  21. [21]

    Grounded text-to-image synthesis with attention refocusing.2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7932–7942, 2024

    Quynh Phung, Songwei Ge, and Jia-Bin Huang. Grounded text-to-image synthesis with attention refocusing.2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7932–7942, 2024

    2024

  22. [22]

    SDXL: Improving latent diffusion models for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  23. [23]

    In- stancediffusion: Instance-level control for image generation.2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6232–6242, 2024

    Xudong Wang, Trevor Darrell, Sai Saketh Rambhatla, Rohit Girdhar, and Ishan Misra. In- stancediffusion: Instance-level control for image generation.2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6232–6242, 2024

    2024

  24. [24]

    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. InProceedings of the IEEE/CVF International Conference on Computer Vision, 2023

    2023

  25. [25]

    Lay- outdiffusion: Controllable diffusion model for layout-to-image generation

    Guangcong Zheng, Xianpan Zhou, Xuewei Li, Zhongang Qi, Ying Shan, and Xi Li. Lay- outdiffusion: Controllable diffusion model for layout-to-image generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22490–22499, June 2023

    2023

  26. [26]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6818–6828, 2024

    2024

  27. [27]

    3DIS: Depth-driven decoupled image synthesis for universal multi-instance generation

    Dewei Zhou, Ji Xie, Zongxin Yang, and Yi Yang. 3DIS: Depth-driven decoupled image synthesis for universal multi-instance generation. InThe Thirteenth International Conference on Learning Representations, 2025. 11

    2025