acceptodds
Under review as a conference paper at ICLR 2027

The Scene, Not the Shot: Measuring Inferred Composition in Text-to-Image Models

Abstract

Photographers stage emotion geometrically: awe by dwarfing a figure against a vast scene, dominance by tilting the camera upward. These conventions are rarely spelled out—no photographer is told to shrink the subject—yet they are encoded in the photographs text-to-image models are trained on. Models follow explicit framing instructions well; what they do by default, when the framing is not specified, has not been measured. We ask whether models infer the composition an intent calls for, through a benchmark of six axes—awe, romance, social group, power, vulnerability, and relative scale—whose prompts name the intent but never how to frame it. Each axis extracts a geometric quantity from the image and scores it against human photographic practice or an established convention, with bootstrap uncertainty throughout. The reference is a range of human practice, not a single correct answer or a recipe for the emotion. Across twelve recent models we find one consistent split, holding within every model: models get the scene right but not the camera. What lies within the scene—the depth of an awe landscape, the spacing of couples and groups—stays near human practice. How the camera captures it does not: the awe subject is framed too large or too uniformly, couples are shot from too far, and although power calls for a low camera angle and vulnerability a high one, every model keeps the camera at eye level in at least seven of ten generations, tilting it barely more often than when no intent is named. Relative scale compresses: small things render too large, vast things too small. No model dominates across axes. Models stage the scene; they do not shoot it.

Primary area:
datasets and benchmarks
Keywords:
text-to-image generation, benchmark, evaluation, image composition, visual emotion, implicit inference

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_fu3NOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors describe a benchmark called GEOCOMP that considers whether text-to-image models infer the composition an emotional or relational intent calls for, when the framing is never named. Each of the 6 axes compares a geometric quantity measured on the image with human photographic practice or an established convention, with bootstrap uncertainty throughout. The authors argue that current models seem to have learned what photographs contain, but not how photographers place the camera to convey an intent; for this, the framing that gives an image its meaning must still be spelled out in the prompt.

Critical strengths

None

Critical weaknesses

None

Questions for authors

I'm wondering how and why the authors chose two image datasets for this study. Would it be possible to say those chosen datasets are the best choice for evaluating how photographers place the camera to convey his/her intent? This is connected to the robustness issue of this study, i.e., dataset dependency of the conclusions: current models seem to have learned what photographs contain, but not how photographers place the camera to convey an intent.

Reviewer_JEvJOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

this paper propose a benchmark GeoComp, that evaluates how good the text-to-image models can infer suitable camera angle, or scene compositions given a prompt with only emotions but no specific scene layout informations. The hypothesis is that photographers usually stage emotion geometrically, and models trained on these visual data should also preserve this. The evaluations covers several axes, including awe, romance and social group, power and vulnerability, and relative scale. The experimental results show that no model dominates across axes.

Critical strengths
  1. Authors evaluated on multiple text-to-image models using their proposed benchmark.
Critical weaknesses
  1. The presentation is not optimal, and very hard to follow. For example, the abstract has two paragraphs which is not conventional, the introduction is very hard to follow, the last sentence of the introduction does not make sense to me as a conclusion.

  2. The motivation of this evaluation is unclear. If the intended application is generating images with an appropriate composition, a natural alternative (which is an established way to use t2i models) is a two-stage pipeline: first augment the raw prompt with suitable geometrics, then let the t2i model generate. If the two-stage pipeline can resolve the problem successfully, the paper does not justify why we should care about the implicit camera angle inference capability of t2i models.

  3. The correlation between the emotion and the implicit camera angle is not objective. The benchmark operationalizes expressive intent through selected camera angles, subject scales, and spatial arrangements. However, an image can plausibly convey power, vulnerability, or awe through combinations of pose, lighting, expression, etc. Furthermore, if the intention is for content creation (as argued in the introduction), limiting certain emotion to certain camera angle constrains the creativity in arts.

  4. The reliability of the geometric evaluators on generated images is insufficiently established in the paper. The evaluation results would be more convincing if the authors show that the used models for automatic evaluation have good accuracy and are generalisable to generated samples across different generative models.

  5. The reference set construction may be biased. The reference construction pipeline is not discussed in the paper. In the appendix, it seems that there is a human annotated category "awe" for example in the chosen dataset, but then authors still conduct a series of filtering steps to get the reference set using the criterias that they defined that should be correlated with "awe". This creates a circular evaluation set up.

Questions for authors

see weaknesses.

Reviewer_o1NpOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies implicit intent communicant and whether T2I models can infer photographic composition correctly. Concretely, the benchmark includes 6 intents: awe, romance, social group, power and vulnerability, and relative scale. It benchmarks 12 most current and sota private and opensource T2I models and propose several geometric & statistical measurement to evaluate whether generated images follow photographic composition conventions associated with the intended emotions (e.g. awe should depict small human in vast landscape).

Critical strengths
  • Interesting evaluation perspective. The paper studies whether models infer photographic composition from implicit intent rather than following explicit geometric instructions. The implicit reasoning and prompt understanding as understudied yet important

  • Broad model evaluation and baseline. The evaluation covers 12 recent models (commerical and open source), with explicit geometric measurements that help identify different compositional behaviors.

Critical weaknesses

[W1] Limited benchmark validity and insufficient evaluation. The benchmark evaluates abstract intents (e.g., awe, romance, power, etc) using a narrow set of geometric/statistical measures. No human perceptual evaluation is provided to validate whether the metrics reflect successful intent realization. Furthermore, the heavily curated human-reference distributions may not represent the broader range of valid compositions. Several benchmark choices also lack sufficient justification, e.g., the 9 camera-angle criteira, and the sources of real-world object dimensions as no referneces seem to be cited. Overall, to validate the paper's contirubtions adequate and complete human evals are required to (1) direclty evaluate intent realization (2) metric effectiveness. Additionally, I'd appreciate if authors could clarify the the dtails of benchmark creation, sources.reasoning of selected thresholds . Additonally the limited human anntoators are not known how they are recuirted? What platofrm are used ? and how quality is ensured?

[W2] Unclear evaluation target across heterogeneous model pipelines. Several evaluated systems are unified multimodal models (e.g, Hunyuan, GPT-image, Nano bananna ,etc) or may incorporate reasoning and prompt rewriting, particularly proprietary models with undisclosed pipelines. The paper does not adequately account for these differences, making it unclear whether the results reflect image-generation capabilities or upstream intent interpretation. This is in fact contradictory to the paper's motivation of evaluating implicit composition inference.

[W3] Poor presentation and reproducibility concerns. Despite few grammatical errors, the paper is extremely difficult to follow and requires substantial effort to understand the text. Additionally, the released scoring code appears inconsistent with some of the equations in the paper -- e.g., the romance distance formula (sec 3.2), where both the numerator and normalization denominator differ from the paper. These discrepancies raise reproducibility concerns.

Questions for authors
  1. Human evaluation for (1) validate whether the proposed geometric metrics correlate with human scores (2) directly evaluate at-least a subset of benchmark can significantly help understanding the contributions of the paper

  2. Since the benchmark is the main contribution. Please provide details on how benchmark is created, beyond mentinong the datasets that images were constrcuted from. e.g., How sensitive are the results to the selection of human-reference images and benchmark thresholds? What are the sources of the real-world object dimensions used for relative-scale evaluation? What references

  3. How do the authors account for differences in reasoning and prompt-rewriting capabilities across evaluated models, particularly proprietary systems with undisclosed inference pipelines?

see details in weaknesses

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.