acceptodds
Under review as a conference paper at ICLR 2027

CroGen: Probing Spatial Control at Extreme Target Densities in Image Generation

Abstract

Advanced image diffusion models generate realistic images and support flexible spatial control. However, we find that spatial correspondence deteriorates as the number of instances we aim to control increases, despite visually convincing outputs. We term this phenomenon spatial constraint overload. In this paper, we explore this phenomenon using dense crowds as a case study and introduce CroGen, a simple and elegant method to alleviate it. Specifically, CroGen combines Proximity Field Conditioning (PFC) and Pixel-Packing Injection (PPI). Our PFC converts target locations into a quantized distance field, providing dense guidance while keeping individual locations explicit. Our PPI retains fine spatial details through pixel packing and injects the packed condition through a single linear projection. We further introduce a dataset of 71K crowd images spanning diverse crowd densities, with point annotations and text prompts. Comprehensive experiments on controllable generation and downstream dense crowd counting show that our CroGen improves spatial consistency without sacrificing image quality. In particular, our method better maintains this consistency as the number of controlled instances grows, demonstrating its effectiveness in alleviating spatial constraint overload. The dataset and model will be released.

Primary area:
generative models
Keywords:
Dense Spatial Control, Controllable Image Generation, Diffusion Models, Spatial Condition Encoding, Crowd Generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_K3yeOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CroGen studies loss of spatial control in dense crowd generation. It combines a quantized nearest-point distance field with pixel-packed condition injection into a diffusion transformer, and introduces a 71K-image dataset. Experiments evaluate point alignment, image quality, and the usefulness of synthetic images for crowd counting.

Critical strengths

The method is simple and addresses a concrete failure mode. The density-stratified results and downstream counting experiment support its practical value: spatial alignment improves without an obvious image-quality penalty. Overall, this is a useful empirical contribution.

Critical weaknesses
  1. The comparison does not establish how the method fares against strong general-purpose generators such as Nano Banana and GPT Image. DenseControl is discussed but also absent from Table 1. These comparisons would better establish the current capability gap; for API models, differences in supported spatial inputs should be made explicit.

  2. The main spatial evaluation uses perturbed mosaics of training condition maps (Sec. 4.1). This is useful as a stress test, but leaves generalization to genuinely held-out crowd layouts less clear. The separate validation-image quality test does not directly resolve that issue.

Questions for authors
  1. Have Nano Banana, GPT Image, or DenseControl been tested on these layouts? If so, please report the results and exact conditioning protocol; otherwise clarify the comparison limitations.
  2. Are spatial Precision/F1 results available on the held-out source test splits, and are the processed training/validation crops separated by original image?
Reviewer_6miiOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Recent diffusion models can accept spatial conditions, but as the number of instances to control grows, the generated objects drift away from their target locations, merge with one another, or vanish altogether, even though the image still looks plausible. The authors call this phenomenon spatial constraint overload and analyze it using dense crowd images as a case study. To address it, they propose Proximity Field Conditioning (PFC), which converts target point locations into a distance field for conditioning, and Pixel-Packing Injection (PPI), which prevents the conditioning information from being washed out. Built on FLUX.1-dev, CroGen generates crowd scenes better than competing methods.

Critical strengths
  • The paper sets up a novel task: image generation that faithfully satisfies conditions for a very large number of people (>100).
  • It proposes a new solution to the washing-out of conditioning information that occurs when many conditions are densely packed into a single location.
  • Although it is not stated explicitly, the tone of the submission suggests that the authors applied their method to the same model to ensure a fair comparison.
  • The ablations are thorough: they cover how the condition map is generated, performance when toggling PFC and PPI, and performance across different ranges of the number of generated people (though I think this last part would be better placed in the main results section).
Critical weaknesses
  • The paper proposes the task of generating a very large number of people and solves it, which I acknowledge. However, the utility of this task is not clearly explained in the paper (it appears to mention only spatial control). As a result, this reviewer is not fully convinced of why this problem needs to be solved, given that it is not generally in demand, and the main text needs additional explanation on this point.
  • It is unclear how many images were used to compute FID. Given that the validation set has 2,341 images, if that is the number used, the FID is an unreliable number. FID is a biased estimator, and producing a reliable value requires generating at least 10k images. I recommend running experiments with a metric such as KID (Kernel Inception Distance) instead of FID.
  • The meaning of CLIP-I here is unclear. CLIP-I is used as an image quality metric, but it only measures the similarity between two images, so it is hard to conclude that the method yields better image quality. I strongly recommend supporting this claim with an aesthetic predictor or another image quality assessment metric.
  • For Spatial Cons., head locations are detected with P2PNet, which itself has a non-trivial error rate. Taking this into account, I recommend running an appropriate statistical test to confirm that CroGen achieves significantly higher performance than FLUX.1 or OmniControl. (In particular, I strongly recommend measuring the error rate as a function of the number of people, since my intuition is that the error grows as the number of people increases.)
  • An analysis of what happens when generating a small number of people also seems necessary. Since supervision is currently given only at head points, does generating few people cause a problem where only heads float around?
  • OmniControl, as the name suggests, supports a wide variety of guidance types. However, this paper validates only the number of people and does not verify other forms of control, so a discussion of, or experiments on, generalizability would be welcome.
  • A paper should always be self-contained and reproducible. This submission lacks the training setup, evaluation measures, and dataset descriptions, in both the main text and the Appendix (which does not exist in the first place), and no code is provided as supplementary material either. This is a very critical problem.
Questions for authors
  • I think a layout is a stronger condition than a point. Is it possible to compare against models or methods that provide guidance through layouts?
  • The crowd size distribution appears to be highly unbalanced. Has this been taken into account?
  • In Pixel-Packing Injection, were other injection schemes considered? Also, doesn’t quantization lose information anyway? To strengthen the submission’s claim, it would be necessary to verify whether existing approaches really lose information and whether alternatives exist, how localization precision changes when τ\tau or KK is varied, and how the method behaves when an operation other than addition is used. (Why can you say it is near-lossless? Is there any theoretical or empirical results that can support that claim?)
Reviewer_m76xOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies a very narrowly defined research problem: how to achieve accurate spatial correspondence when image generators are conditioned on increasingly dense target layouts. The authors term the degradation of this correspondence “spatial constraint overload.” Using crowd generation as a case study, the paper proposes CroGen, which combines a quantized signed distance map (SDF) with pixel packing and a linear projection that adds spatial conditions to the DiT input embeddings. The authors construct approximately 71K samples from existing crowd datasets and evaluate spatial alignment, image quality, and downstream crowd counting. The reported results show improved positional accuracy, particularly at high target counts, while maintaining image quality.

Critical strengths
  • Simplicity and practical utility.

The major advantage of this paper is at the proposed method is straightforward and easy to understand.

The conditioning mechanism uses pixel rearrangement and a single linear projection, avoiding a separate ControlNet-style control branch. The reported improvements over OminiControl, particularly in dense layouts, support its practical value for generating crowd images with more accurate head locations and more reliable point annotations. This combination of architectural simplicity and measurable task-specific benefits is the paper’s strongest contribution.

  • Clear empirical evaluation within the proposed crowded people generation domain.

The paper provides an extensive evaluation of crowd generation, including comparisons with multiple baselines, component ablations, and performance analysis across different target densities. It assesses both spatial alignment and image quality, and further evaluates the utility of the generated data through downstream crowd counting on real images. These complementary evaluations provide substantial empirical support for the method’s effectiveness within the studied setting.

Critical weaknesses
  • Limited technical novelty.

The proposed method combines a distance-field representation with pixel rearrangement, linear projection, and additive conditioning. Although this combination is sensible, the manuscript does not convincingly establish a substantial methodological advance beyond a straightforward design for preserving spatial information. Improved localization is consistent with providing more accessible position cues, but the paper offers limited additional insight into when and why this particular combination is necessary. The major concern is therefore the depth of the contribution, rather than simplicity itself.

  • Empirical evaluation is narrowed in the target domain.

All experiments concern crowd generation or downstream crowd counting, leaving the applicability of the proposed representation and injection mechanism to other dense spatial-control settings unestablished. The image-quality results primarily support maintaining quality: for example, the full method changes FID from 48.3 to 47.7 and CLIP-I from 0.761 to 0.779 relative to the FLUX.1 baseline. These results are useful, but provide limited evidence of benefits beyond the specific localization objective. A focused application can constitute a strong contribution; here, however, the narrow empirical scope does not sufficiently compensate for the limited methodological novelty. On this basis, my recommendation is reject.

Questions for authors

See the Weaknesses for more details.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.