cs.CVNov 5, 2026
SaveDo You Miss the Nights We Spent Crafting Figures? Rethinking Teaser and Framework Figures in the Age of Generative AI
Organizations: Nanyang Technological University, Singapore
Abstract
Scientific figures give research ideas a visual form through choices of abstraction, composition, and emphasis. Generative AI offers new ways to realize these choices, raising questions about explanatory clarity and purposeful visual expression. We examine these questions through a visual retrospective of 371 motivation and framework figures from 200 papers. The corpus comprises a quota-based main sample of 160 papers spanning four temporal periods and a separate purposive style-contrast sample of 40 papers. A single multimodal judge evaluates six dimensions of communication and design, with three repeated judgments per figure. We attempt source-constrained revisions of all 371 figures, of which 369 pass the recorded fidelity checks. Original-figure ratings show no uniform decline across periods. Revisions increase mean clarity ratings by 0.428 points and decrease aesthetic coherence ratings by 0.111 points on a five-point scale. Gains in the six-dimension average are larger for main-sample motivation figures than for frameworks. Case analyses show that added explanation can clarify comparisons while leaving typographic hierarchy and visual economy unresolved. These model-based findings support researcher-led design, with AI assisting exploration and authors retaining responsibility for explanatory purpose, composition, and scientific relationships. The accompanying atlas and evaluation records are available at https://ntu-ai4x.github.io/Rethinking_Figures/index_en.html.
Figures & tables
Figure 1: Distinctive visual designs in U-Net, CLIP, and the Transformer, reproduced from Figure 1 of the respective papers [ 1 , 3 , 2 ] . These illustrative examples are outside the analytical corpus.
Figure 2: Main-sample design across temporal stages, candidate source pools, and recurring research topics. Source labels identify the search scope, and release milestones indicate tool availability. Timeline lengths are schematic.
Figure 3: Two collection paths produce 200 papers and 371 figures. A: Role and version checks support quota-based selection of paired motivation (M) and framework (F) figures. B: Purposive selection retains eligible M and/or F figures. Candidate pools overlap. In the contrast pool, 47 PDFs underwent visual inspection.
Figure 4: Corpus composition. A: Main-sample allocation and S4 eligibility tiers. B: Fixed PDF sources, distinct from publication venues. C: OpenReview submission areas of the contrast sample, with shortened metadata labels. D: Motivation (M) and framework (F) figure counts. The contrast sample is excluded from temporal stage averages.
| Year of verified figure version | ||||||||||
| Stage | 2018 | 2019 | 2020 | 2021 | 2022 | 2023 | 2024 | 2025 | 2026 | Total |
| S1 | 3 | 3 | 1 | 21 | 52 | – | – | – | – | 80 |
| S2 | – | – | – | – | – | 3 | 35 | 42 | – | 80 |
| S3 | – | – | – | – | – | – | – | 20 | 60 | 80 |
| S4 | – | – | – | – | – | – | – | – | 80 | 80 |
| Total | 3 | 3 | 1 | 21 | 52 | 3 | 35 | 62 | 140 | 320 |
Table 1: Temporal distribution of verified figure versions in the main sample. Entries count figures, and dashes denote zero.
| Dimension | Score 1 | Score 3 | Score 5 |
|---|---|---|---|
| Communicative Clarity | The central message is unclear, and key objects or relationships require substantial inference. | The main problem or process is identifiable, but important relationships or emphasis remain unclear. | The intended problem or contrast, or the main inputs, operations, relationships, and outputs, are readily understood. |
| Organization & Readability | Confusing hierarchy, crowding, or illegible key text obstructs the reading sequence. | Most information can be followed, but grouping, density, hierarchy, or text requires noticeable effort. | Reading order and hierarchy are clear, grouping manages complexity, and key labels are legible. |
| Graphical Consistency | Conflicting or unexplained visual encodings make important relationships ambiguous. | Encodings are mostly consistent, with some ambiguous labels, directions, connections, or repeated elements. | Visual encodings are consistent and traceable, with necessary labels or legends clarifying relationships. |
| Aesthetic Coherence | Clashing typography, colors, spacing, or styles compete with the message. | The composition is broadly coherent but has uneven alignment, spacing, emphasis, or balance. | Color, typography, spacing, alignment, and balance jointly support the information. |
| Purposeful Visual Inventiveness | Generic arrangement or ornament contributes little problem-specific understanding. | A deliberate comparison, arrangement, example, or explanatory device helps, but much remains generic. | A well-chosen explanatory device provides a particularly effective account of the concept. |
| Design Specificity | Replacing the labels would allow the design to represent many unrelated problems. | Some elements are tailored to the problem, while much of the design remains interchangeable. | The elements and their arrangement express distinctive problem-specific objects, relationships, or operations. |
Table 2: Operational dimensions and compact scoring anchors. Scores of 2 and 4 indicate performance between adjacent anchors. The same frozen rubric is used for original and revised figures.
Figure 5: Original-figure profiles across stages and roles. Cells show means of per-figure medians from three judgments, and panel labels report figure counts. The main sample retains equal topic quotas, and the purposive contrast sample is displayed separately. All panels share the same 1–5 color scale.
| Dimension | Original | Revised | 95% interval | Up / Same / Down | |
|---|---|---|---|---|---|
| Communicative clarity | 3.629 | 4.057 | +0.428 | [+0.367, +0.492] | 142 / 220 / 7 |
| Organization and readability | 3.973 | 3.951 | -0.022 | [-0.084, +0.040] | 51 / 257 / 61 |
| Graphical consistency | 3.949 | 4.060 | +0.111 | [+0.056, +0.168] | 69 / 268 / 32 |
| Aesthetic coherence | 3.591 | 3.480 | -0.111 | [-0.170, -0.054] | 38 / 252 / 79 |
| Purposeful visual inventiveness | 3.450 | 3.501 | +0.051 | [+0.014, +0.089] | 35 / 317 / 17 |
| Design specificity | 4.461 | 4.623 | +0.163 | [+0.114, +0.212] | 66 / 293 / 10 |
Table 3: Paired dimension scores for 369 revisions that pass the recorded fidelity checks. Values are means of figure-level medians from three judgments, and denotes revised minus original. Intervals are 95% paper-cluster bootstrap percentile intervals based on 5,000 resamples with seed 20261020. Direction counts refer to figure-level median changes.
Figure 6: Paired outcomes for 369 revisions that pass the recorded fidelity checks. Left: counts of increased, unchanged, and decreased figure-level medians. Right: mean changes and 95% paper-cluster bootstrap percentile intervals.
| Sample | Role | Original | Revised | 95% interval | Up / Same / Down | ||
|---|---|---|---|---|---|---|---|
| Main | M | 160 | 3.818 | 4.059 | +0.242 | [+0.186, +0.298] | 107 / 29 / 24 |
| Main | F | 159 | 3.824 | 3.842 | +0.018 | [-0.018, +0.055] | 57 / 54 / 48 |
| Style contrast | M | 10 | 4.067 | 4.133 | +0.067 | [-0.083, +0.233] | 4 / 3 / 3 |
| Style contrast | F | 40 | 3.954 | 3.854 | -0.100 | [-0.167, -0.033] | 5 / 16 / 19 |
Table 4: Paired six-dimension averages by sample and figure role. M and F denote motivation and framework figures, and is the number of accepted original–revision pairs. Intervals are 95% paper-cluster bootstrap percentile intervals within each row. Direction counts refer to changes in the post hoc six-dimension average.
| Motivation | Framework | |||||||
|---|---|---|---|---|---|---|---|---|
| Topic | ||||||||
| Object detection | 40 | +0.900 | +0.125 | +0.367 | 40 | +0.150 | -0.150 | +0.046 |
| Image segmentation | 40 | +0.600 | -0.050 | +0.146 | 39 | +0.205 | -0.128 | +0.013 |
| Image restoration | 40 | +0.850 | -0.100 | +0.225 | 40 | +0.150 | -0.125 | +0.021 |
| 3D vision | 40 | +0.850 | -0.100 | +0.229 | 40 | +0.100 | -0.250 | -0.008 |
Table 5: Main-sample revision changes by topic and figure role, pooling S1–S4. and denote mean changes in clarity and aesthetic coherence, while denotes the change in the post hoc six-dimension average. Counts refer to accepted figure pairs. One segmentation framework failed fidelity review, leaving 39 pairs in that cell and 40 in every other cell.
| Pattern | Figures | Percentage |
|---|---|---|
| Gains without losses | 153 | 41.5% |
| Losses without gains | 80 | 21.7% |
| Mixed gains and losses | 57 | 15.4% |
| All six dimensions unchanged | 79 | 21.4% |
| Total | 369 | 100.0% |
Table 6: Mutually exclusive patterns of dimension-level change across 369 accepted revisions. A gain or loss denotes an increase or decrease in a figure’s median score on at least one dimension. The unchanged category requires all six medians to remain identical. Percentages use 369 figures as the denominator.
| Original | Revised | ||||
|---|---|---|---|---|---|
| Case | Score vector | Mean | Score vector | Mean | |
| 163 F: AccessWorld [ 53 ] | 4.83 | 5.00 | +0.17 | ||
| 128 M: Over-segmentation [ 54 ] | 2.67 | 4.33 | +1.67 | ||
| 136 F: ReTax-Seg [ 55 ] | 3.00 | 3.00 | 0.00 | ||
| 195 F: Stride [ 56 ] | 3.50 | 3.00 | -0.50 | ||
Table 7: Absolute score extrema selected by the post hoc six-dimension mean. Entry 163 F is the unique maximum within the original set and within the accepted-revision set. Entry 128 M is the unique original minimum, while 136 F and 195 F tie for the accepted-revision minimum. Vectors follow , and means and changes are computed before rounding. These ranks concern model-rated figures, not the scientific quality of their papers.
Figure 7: Absolute original-score extrema and their revisions. AccessWorld (163 F) is the unique original maximum, and the over-segmentation example (128 M) is the unique original minimum. Both were identified after scoring using the six-median average. Images are reproduced in full, with three-request median scores in C/O/G/A/I/S order. Original artwork is credited to the cited papers, and revisions are study-generated derivatives.
Figure 8: The two accepted revisions tied for the lowest six-score mean (3.00). ReTax-Seg (136 F) retains all six medians, while Stride (195 F) loses one point each in organization, aesthetic coherence, and inventiveness. Selection is post hoc and illustrative. Full images and three-request median scores are shown. Original artwork is credited to the cited papers.
Figure 9: Typographic changes in two motivation figures (062 M and 198 M). Added headings repeat retained labels in Complexity Experts. A conventional title and matching heading strips replace the integrated title treatment in VEvo-Harness. C, O, G, A, I, and S denote clarity, organization, graphical consistency, aesthetic coherence, inventiveness, and specificity. Scores are three-request medians, and changes are computed before display rounding. Images are shown in full. Original artwork is credited to the cited papers, and revisions are study-generated derivatives.
Figure 10: Added explanatory layers in BKBAgent (167 M). Three connected explanation boxes and a summary sentence extend the original composition. Clarity and organization remain unchanged, while aesthetic coherence and inventiveness each decrease by one point. Both images are shown in full, with three-request median scores in C/O/G/A/I/S order. Original artwork is credited to the cited paper, and the revision is a study-generated derivative.
Figure 11: Repeated content in two framework revisions (147 F and 195 F). Enlarged details duplicate regions retained in the complete overviews, repeating existing labels and pictograms. Organization decreases in both cases. Images are fitted individually without cropping, preserving their differing canvas proportions. Scores are three-request medians in C/O/G/A/I/S order. Original artwork is credited to the cited papers, and revisions are study-generated derivatives.
Figure 12: Contrasting annotation outcomes in motivation figures (100 M and 127 M). SegEarth-OV3 gains in five dimensions while retaining its aesthetic score. PULSE gains one clarity point and loses one aesthetic point, leaving the six-score mean unchanged. Images are shown in full, with three-request median scores in C/O/G/A/I/S order. Original artwork is credited to the cited papers, and revisions are study-generated derivatives.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
| ID | Cohort | Paper (clickable title and citation) | M/F |
|---|---|---|---|
| 001 | S1 | A Dual Weighting Label Assignment Scheme for Object Detection [ 13 ] | M/F |
| 002 | S1 | Cross Domain Object Detection by Target-Perceived Dual Branch Distillation [ 61 ] | M/F |
| 003 | S1 | Dense Learning Based Semi-Supervised Object Detection [ 62 ] | M/F |
| 004 | S1 | Localization Distillation for Dense Object Detection [ 14 ] | M/F |
| 005 | S1 | Not All Labels Are Equal: Rationalizing the Labeling Costs for Training Object Detection [ 63 ] | M/F |
| 006 | S1 | Progressive End-to-End Object Detection in Crowded Scenes [ 64 ] | M/F |
More in this topic
Scientific Visualization
Scientific visualization turns complex scientific data into visual representations for analysis, exploration, and communication.
See all 54 papers