How to Make Your Multi-Image Posts Popular? An Approach to Enhanced Grid for Nine Images on Social Media

Binhui Wang; Jianchao Zhang; Liangxu Wang; Qi Xi; Shulin Li; Shunye Tang; Yan Zhang; Yiru Niu; Zhiqi Gao; Zibo Zhang

REVIEW 4 major objections 6 minor 11 references

How to Make Your Multi-Image Posts Popular? An Approach to Enhanced Grid for Nine Images on Social Media

T0 review · 4 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read Putting a post's most aesthetically rated image in the center of a nine-grid wins the most user votes.

desk verdict Useful four-way comparison of nine-grid arrangements, but the headline preference is statistically fragile and the design confounds ordering with position; worth refereeing as a pilot study. read the letter →

arxiv 2502.03709 v2 pith:7NTS52JZ submitted 2025-02-06 cs.HC

Qi Xi , Shulin Li , Zhiqi Gao , Zibo Zhang , Shunye Tang , Jianchao Zhang , Liangxu Wang , Yiru Niu

show 2 more authors

Yan Zhang Binhui Wang

This is my paper · ORCID

classification cs.HC

keywords nine-gridlayoutimagearrangementaestheticqualitypopularitycenterprioritizationuserpreferencesocialmediaforced-choicestudy

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 asks whether the order of nine images in a three-by-three grid changes how much people like the post, and which arrangement rule people prefer. It builds four candidate layouts from two single-image scorers (aesthetic quality and intrinsic popularity) and two ordering principles (left-to-right sequential and center-first), then asks 45 participants to pick their favorite layout in 50 forced-choice comparisons each. The aesthetic-quality layout with the best image in the center collected the most votes (628 of 2,250), and the authors conclude that arranging by aesthetic quality and prioritizing the center yields more favorable user evaluations. If that holds, content creators can improve the perceived quality of multi-image posts without changing any image content, just by reordering thumbnails.

What carries the argument

The central objects are two per-image scoring models and two ordering rules. NIMA scores each thumbnail's aesthetic and technical quality; I2PA scores its intrinsic viral potential. The ordering rules are sequential (reading order from top-left to bottom-right, best first) and center prioritization (best image in the central cell, next four in the corners, rest on edges). Crossing the two scorers with the two rules yields the four nine-grid layouts compared in the forced-choice test.

What would settle it

Rerun the forced-choice study with layouts whose center image is the one human raters pick as best rather than the NIMA-picked best; if the aesthetic-center scheme then loses its advantage, the reported effect depends on the model's ranking, not on the center-prioritization principle itself.

Watch

Extended reading notes

Core claim

The paper claims that, given the same nine images, the arrangement that places the highest-scoring image in the center and distributes the next four by score into the corners, with all ordering decided by per-image aesthetic quality scores (NIMA) rather than predicted viral-popularity scores (I2PA), receives the most favorable user evaluations. In a forced-choice study with 45 participants and 50 trials each (2,250 votes), this aesthetic center-prioritization scheme collected 628 votes, ahead of aesthetic sequential (599), content center (585), and content sequential (438). The authors conclude that arranging images based on aesthetic quality and prioritizing central positions leads to more favorable user evaluations.

Load-bearing premise

The comparison assumes that NIMA and I2PA scores rank the nine thumbnails the same way a human viewer would, even though both models were trained on single images and are never validated on nine-grid crops.

Editorial extensions

If this is right

A content creator who puts the highest-quality image in the center and fills the corners with the next four best images can expect more favorable evaluations than a chronological or content-only ordering.
Aesthetic quality, as measured by a single-image model, predicts user preference for grid arrangement better than predicted popularity or virality scores.
The serial-position view (first and last positions matter) is less predictive of user preference than the visual-attention view (center matters) for nine-grid layouts.
An automated arrangement tool can be built on these two scores and two rules, since the rearrangements are computable without human input per post.

Reading between the lines

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

If the result generalizes, the benefit may transfer to other multi-image grids, such as 3x3 mosaics in stories, profile layouts, or product galleries, and may also hold when the best image is chosen by human raters rather than by NIMA.
The forced-choice test measures immediate preference, not social-media engagement; actual likes, comments, and shares could behave differently, which the paper does not measure.
A direct test of the mechanism would be to let the same crowd rank the nine images themselves and compare layouts built on their own rankings against layouts built on model rankings; if model-built layouts still win, the conclusion is about layout geometry rather than score accuracy.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Quick take on arXiv:2502.03709.

The genuinely new thing here is the four-cell comparison: two model-based orderings (NIMA aesthetic scores and I2PA popularity scores) crossed with two spatial principles (sequential reading order and center-prioritized placement), tested by forced choice on the same nine-image sets. That specific comparison does not appear in the prior work they cite, and the data collection is transparent and easy to follow. The formative survey and interviews are honest hypothesis-generation, not rigged evidence: user preferences were collected before the schemes were defined, and the model scores are external, so there is no circular fitting of outcome to predictor.

The soft spot is the headline result. Table I gives 628, 599, 585, and 438 votes out of 2250. The difference between 628 and 599 is not significant, and 628 vs 585 is weak; the only robust contrast is that content-sequential (438) is worse than the other three. So the abstract's claim that "layout schemes based on aesthetic quality outperformed others" is not supported by the data. The apparent aesthetic/center-priority advantage is driven by one low cell.

Two design issues amplify the problem. First, the 2250 votes are clustered: 45 participants answered 50 questions about 250 image sets, yet there are no significance tests, confidence intervals, or participant/image-set random effects. Second, the ordering principle is confounded with position assignment: center-priority puts the top image in P5 and the next four in the corners; sequential puts ranks 1-5 in P1-P5. Any preference for center-priority could come from the corners being favored, which their own formative survey found. There is also no baseline against the original arrangements, and the NIMA/I2PA models are applied to cropped thumbnails without any validation that their rankings match human rankings in this context.

That sounds like a takedown, but it is not. The paper is a reasonable pilot study with a clear design, an honest write-up, and practical value for content creators if the effects hold. The problem is the strength of the conclusion, not the research question. If the authors reframe the result as a hypothesis, add mixed-effects modeling or at least proper significance tests, include the original layout as a baseline, and release the stimulus set and vote data, the paper would be genuinely useful.

For peer review: yes, a serious editor should send it to referees. The topic is timely, the comparison is new, and the methodological gaps are fixable. I would not cite it yet for the central claim, but I would bring it to a reading group as an example of how to design a clean four-way comparison and then over-interpret the top cell.

Referee Report

4 major / 6 minor

Summary. The paper investigates how the arrangement of nine images in a three-by-three grid affects user preference and perceived popularity of multi-image social media posts. It uses two pretrained models, NIMA for aesthetic quality and I2PA for intrinsic visual popularity, to rank the nine images of each of 250 Weibo posts, and constructs four layout schemes: aesthetic-sequential, aesthetic-center-prioritization, I2PA-sequential, and I2PA-center-prioritization. Participants (N=45) completed forced-choice questionnaires selecting their preferred layout among the four for each image set, yielding 2,250 total votes. The reported vote counts (Table I) are 628, 599, 585, and 438 for aesthetic-center, aesthetic-sequential, I2PA-center, and I2PA-sequential, respectively. The paper concludes that arranging images by aesthetic quality and prioritizing the central position yields the most favorable user evaluations.

Significance. If the central claim were well supported, the paper would offer directly actionable guidance for content creators and social media managers, and it would contribute empirical evidence on whether serial-position effects or central visual attention better predict preference in multi-image grids. The study has notable strengths: it uses real social-media image sets (1,901 crawled posts, 250 selected), derives the layout schemes from two explicit theoretical positions, and collects preference data through a concrete forced-choice method rather than relying solely on model predictions. The use of established pretrained models (NIMA, I2PA) is also a reasonable starting point. However, the statistical basis for the headline result is currently thin, and the experimental design contains a confound that prevents the paper from isolating the effect it claims to establish. The contribution is therefore promising but requires substantial strengthening before the conclusions can be accepted.

major comments (4)

[Section V.B, Table I] The conclusion that the aesthetic center-prioritization scheme 'yielded the best results' is not supported by the reported aggregate counts. The leading vote counts are 628, 599, and 585 out of 2,250; pairwise two-proportion tests give p ≈ 0.41 for 628 vs. 599 and p ≈ 0.22 for 628 vs. 585, and the absence of participant-level or image-set-level clustering corrections makes the effective sample size smaller than 2,250. The paper reports no significance test, confidence interval, or mixed-effects model, so the only robust contrast is that the I2PA-sequential scheme (438) is disfavored; the three other schemes are statistically indistinguishable.
[Section IV.C and V.B] The comparison between 'center prioritization' and 'sequential' is confounded with which grid positions receive the top five images. In the center-prioritization scheme the top image occupies P5 and the next four images occupy the four corners, whereas in the sequential scheme ranks 1–5 are placed in P1–P5. The formative study (§III.A) itself reports that corners (P1, P3, P7, P9) are preferred over edge positions, so the relative success of the center-prioritization condition may reflect the corner placement of images 2–5 rather than a specific effect of the central position. An arrangement that varies the two principles independently is needed to disambiguate the cause of the preference.
[Section IV.A–IV.B and V.A] The NIMA and I2PA models are used to rank thumbnails, but these models were trained on single-image tasks and are applied here to 300×300 crops without any validation of their rank ordering on this material. The entire comparison among the four schemes presupposes that the model scores correctly identify which images are 'best' and 'worst'; if the models mis-rank images, the resulting layouts are effectively arbitrary permutations and the vote counts reflect noise rather than arrangement principles. The paper should provide evidence such as correlation of model scores with human ratings on a sample of the stimuli that the model-based rankings are trustworthy for thumbnail crops.
[Section V.B] The forced-choice task compares only the four rearranged schemes; the original arrangement of each post is not presented as an option. Because the claimed practical benefit is that the proposed layouts improve on what creators currently do, the absence of the original (or a random) baseline makes it impossible to attribute the observed preferences to improvement over the status quo. The separate pairwise comparison in §III.A (rearranged vs. original) uses a different task and sample, and cannot be imported into the main experiment's conclusions.

minor comments (6)

[Section III.A] The text refers to 'Figure 2' when describing the pairwise image-pair comparison, but the caption for Figure 3 identifies it as the image-pair questionnaire; cross-references should be corrected.
[Section IV.A.2] The statement 'Scores were ranging from -5 to 5' should be clarified with the exact output range of the I2PA model and any scaling applied before ranking; the original I2PA model does not necessarily produce scores in that range.
[Section V.A] The phrase '1,901 nine-gram charts' should read 'nine-grid layouts' or 'nine-grid posts'; the same terminology issue appears in Section II.B.
[Abstract and Section VI.A] The term 'OGMI' is introduced in Section I but not used later, and Section VI.A describes a 'Generative tool' that is not detailed or evaluated; either provide specifics or place the tool description in clearly labeled future work.
[Abstract] The phrase 'experiential-centered evaluation' should be 'experience-centered evaluation' or 'user experience evaluation'.
[References] The in-text citation style is inconsistent (e.g., 'Keyan Ding et al [1]' and 'Hossein Taleb [2]'); please unify the formatting to the journal's style.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the evaluation is empirical and self-contained.

full rationale

The paper's derivation chain is not circular. The four arrangement schemes are produced by applying two external, pre-trained scoring models (NIMA and I2PA) to rank images, and then placing them according to two pre-specified ordering rules (sequential and center-prioritization). The outcome measure is independent human forced-choice voting among the four schemes for the same image set. No parameter is fitted to the vote data, and the scheme labels ('aesthetic' vs. 'content', 'sequential' vs. 'center') are not defined in terms of the vote outcome. The formative survey informed which ordering rules to test, but that is hypothesis generation rather than circular reasoning: the later user experiment is a separate empirical test with 45 new participants and 250 image sets. The paper's references to NIMA and I2PA are to external prior work, and no load-bearing self-citation or imported uniqueness theorem appears. The statistical fragility of the headline result (e.g., the closeness of 628 vs. 599 vs. 585 votes) and the confound between ordering principle and position assignment are validity concerns, not circularity. Under the stated criteria, the paper's central claim is self-contained and evidence-based, so the circularity score is 0.

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

The central claim rests on no new fitted parameters. The four layout schemes are generated by external, pre-trained models (NIMA and I2PA) and by two hand-chosen ordering rules. The key assumptions are domain assumptions about the transferability of the models and the representativeness of the dataset. No new entities are introduced.

assumptions (4)

domain assumption The serial-position effect and central visual attention distribution apply to nine-grid image layouts.
Section II.C adopts these two cognitive psychology theories as the basis for designing the sequential and center-prioritization ordering schemes.
domain assumption NIMA and I2PA model scores reflect human judgments of aesthetic quality and content popularity for nine-grid thumbnails.
Sections IV.A-B apply the pre-trained models without validation on nine-grid thumbnail crops; Section IV.C uses the scores to generate all four layouts tested in the main experiment.
domain assumption High-follower Weibo users post higher-quality nine-grid layouts, making their posts a representative source.
Section V.A selects users with more than 2,000 followers to reduce low-quality layouts, but no evidence supports this filtering as representative.
domain assumption A forced-choice preference among four rearranged layouts measures arrangement's effect on post popularity.
Section V.B uses forced-choice votes as the outcome variable; no real engagement metrics (likes, reposts) are collected, so the link to popularity is assumed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of How to Make Your Multi-Image Posts Popular? An Approach to Enhanced Grid for Nine Images on Social Media." pith.science (2026). https://pith.science/paper/7NTS52JZ

@misc{pith2026250203709,
  author       = {Pith},
  title        = {Pith review of: How to Make Your Multi-Image Posts Popular? An Approach to Enhanced Grid for Nine Images on Social Media},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7NTS52JZ}},
  note         = {Machine review of arXiv:2502.03709}
}

read the original abstract

The nine-grid layout is commonly used for multi-image posts, arranging nine images in a tic-tac-toe board. This layout effectively presents content within limited space. Moreover, due to the numerous possible arrangements within the nine-image grid, the optimal arrangement that yields the highest level of attractiveness remains unknown. Our study investigates how the arrangement of images within a nine-grid layout affects the overall popularity of the image set, aiming to explore alignment schemes more aligned with user preferences. Based on survey results regarding user preferences in image arrangement, we have identified two ordering sequences that are widely recognized: sequential order and center prioritization, considering both image visual content and aesthetic quality as alignment metrics, resulting in four layout schemes. Finally, we recruited participants to annotate various layout schemes of the same set of images. Our experience-centered evaluation indicates that layout schemes based on aesthetic quality outperformed others. This research yields empirical evidence supporting the optimization of the nine-grid layout for multi-image posts, thereby furnishing content creators with valuable insights to enhance both attractiveness and user experience.

Figures

Figures reproduced from arXiv: 2502.03709 by the authors.

**Figure 2.** Vote on the position for the best image [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** A set of image pairs from the questionnaire [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Illustration of position 1-9 in a nine-grid format. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: Examples of image rearrangements. arrangement, resulting in four distinct schemes: AestheticsQuality Model-based sequential arrangement, AestheticsQuality Model-based center prioritization arrangement, Intrinsic Image Popularity Assessment Model-based sequential ar…

**Figure 6.** Figure 6: Subjects were asked to choose their favorite of four [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

11 extracted references · 9 canonical work pages

[1]

K. Ding, K. Ma, and S. Wang, ``Intrinsic image popularity assessment,'' in Proceedings of the 27th ACM international conference on multimedia , pp. 1979--1987, 2019

work page 1979
[2]

Talebi and P

H. Talebi and P. Milanfar, ``Nima: Neural image assessment,'' IEEE transactions on image processing , vol. 27, no. 8, pp. 3998--4011, 2018

work page 2018
[3]

Ma and X

X. Ma and X. Meng, ``Image position and layout effects on user engagement of multi-image tweets,'' Proceedings of the Association for Information Science and Technology , vol. 58, no. 1, pp. 490--494, 2021

work page 2021
[4]

X. Ma, Z. He, and Y. Cao, ``What do i suggest you focus on in my photo story? the effect of user personality on the position significance of jiugong grid images in microblog,'' Cyberpsychology, behavior, and social networking , vol. 26, no. 1, pp. 35--41, 2023

work page 2023
[5]

A. M. Colman, A dictionary of psychology . Oxford University Press, USA, 2015

work page 2015
[6]

U ber das ged \

H. Ebbinghaus, \"U ber das ged \"a chtnis: untersuchungen zur experimentellen psychologie . Duncker & Humblot, 1885

work page
[7]

Glanzer and A

M. Glanzer and A. R. Cunitz, ``Two storage mechanisms in free recall,'' Journal of verbal learning and verbal behavior , vol. 5, no. 4, pp. 351--360, 1966

work page 1966
[8]

Khosla, A

A. Khosla, A. Das Sarma, and R. Hamid, ``What makes an image popular?,'' in Proceedings of the 23rd international conference on World wide web , pp. 867--876, 2014

work page 2014

Show all 11 references

[9]

Hessel, L

J. Hessel, L. Lee, and D. Mimno, ``Cats and captions vs. creators and the clock: Comparing multimodal content to context in predicting relative popularity,'' in Proceedings of the 26th international conference on world wide web , pp. 927--936, 2017

2017
[10]

K. He, X. Zhang, S. Ren, and J. Sun, ``Deep residual learning for image recognition,'' in Proceedings of the IEEE conference on computer vision and pattern recognition , pp. 770--778, 2016

2016
[11]

w ) SdWlO W?N؟? IziF 1n pD`δ_( / 0!+y ӗ

11em plus .33em minus .07em 4000 4000 100 4000 4000 500 `\.=1000 = #1 \@IEEEnotcompsoconly \@IEEEcompsoconly #1 * [1] 0pt [0pt][0pt] #1 * [1] 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEcompsocnotconfonly \@IEEEauthorblockAstyle \@IEEEcompsocnotconfonly \@IEEEco...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [1]

K. Ding, K. Ma, and S. Wang, ``Intrinsic image popularity assessment,'' in Proceedings of the 27th ACM international conference on multimedia , pp. 1979--1987, 2019

work page 1979

[2] [2]

Talebi and P

H. Talebi and P. Milanfar, ``Nima: Neural image assessment,'' IEEE transactions on image processing , vol. 27, no. 8, pp. 3998--4011, 2018

work page 2018

[3] [3]

Ma and X

X. Ma and X. Meng, ``Image position and layout effects on user engagement of multi-image tweets,'' Proceedings of the Association for Information Science and Technology , vol. 58, no. 1, pp. 490--494, 2021

work page 2021

[4] [4]

X. Ma, Z. He, and Y. Cao, ``What do i suggest you focus on in my photo story? the effect of user personality on the position significance of jiugong grid images in microblog,'' Cyberpsychology, behavior, and social networking , vol. 26, no. 1, pp. 35--41, 2023

work page 2023

[5] [5]

A. M. Colman, A dictionary of psychology . Oxford University Press, USA, 2015

work page 2015

[6] [6]

U ber das ged \

H. Ebbinghaus, \"U ber das ged \"a chtnis: untersuchungen zur experimentellen psychologie . Duncker & Humblot, 1885

work page

[7] [7]

Glanzer and A

M. Glanzer and A. R. Cunitz, ``Two storage mechanisms in free recall,'' Journal of verbal learning and verbal behavior , vol. 5, no. 4, pp. 351--360, 1966

work page 1966

[8] [8]

Khosla, A

A. Khosla, A. Das Sarma, and R. Hamid, ``What makes an image popular?,'' in Proceedings of the 23rd international conference on World wide web , pp. 867--876, 2014

work page 2014