Making Hero Figures

What Figure 1 should show, with 24 first figures from top papers and what to borrow from each.

The hero figure is Figure 1: the picture a reviewer sees before reading the introduction. In about ten seconds it should answer three questions.

  1. Problem What is hard or missing today?
  2. Idea What is the one thing that changes that?
  3. Payoff What becomes possible, and by how much?

It is a visual abstract, not the architecture diagram: for that, see Making method figures. This page shows 24 first figures from ICML, NeurIPS, ICLR, CVPR, ICCV, ECCV, KDD, AAAI and SIGGRAPH, and what to borrow from each. It is part of Writing Papers.

Choose what leads

Ask which sentence of the abstract a reviewer should repeat to the other reviewers. Draw that sentence.

Your main claim Lead with Layout See
“Everyone has missed problem X.” problem A plot of the failure Informer, ResNet
“A small change to a known method matters.” idea Old and new, side by side Chain-of-Thought, Swin Transformer
“One simple idea solves X.” idea One example in, the idea, the result out CLIP, MAE, LIME, LLMTime
“One method handles many settings.” idea Many inputs, one model, many outputs AnySat, pix2pix
“We do X better, faster or cheaper.” payoff Methods side by side, numbers on the picture 3D Gaussian Splatting, EfficientNet
“We can now do X, which nobody could.” payoff A gallery of real outputs CycleGAN, ImageBind, GeoChat
“Here is a dataset or a benchmark.” scope Real samples and counts, or one item from input to score SatlasPretrain, SWE-bench, CheXpert
“We make several contributions.” scope One panel per contribution Segment Anything

Our default for an applied paper: three panels, problem then idea then payoff, on one real example, with one number. Give the most space to your strongest panel.

Lead with the problem

Use this when the gap itself is your news.

Informer · AAAI 2021 · paper

Two small plots: a short and a long forecast of one series, and an LSTM whose error rises and speed falls as the forecast lengthens.

problem · plot · page 1, one column

Its caption opens: “(a) LSTF can cover an extended period than the short sequence predictions, making vital distinction in policy-planning and investment-protecting.”

It measures an existing model breaking as the forecast gets longer. The new method is not even shown.

Borrow: If a reader may doubt the gap, plot the baseline failing.

Figure 1, Zhou et al. © the authors.

ResNet · CVPR 2016 · paper

Two plots of error against training iterations. In both, the 56-layer curve lies above the 20-layer curve.

problem · plot · page 1, one column

Its caption opens: “Training error (left) and test error (right) on CIFAR-10 with 20-layer and 56-layer “plain” networks.”

Two lines per plot, named on the lines. The deeper network is worse, which nobody expects.

Borrow: A surprising measurement is a complete Figure 1.

Figure 1, He et al. © the authors.

Chain-of-Thought · NeurIPS 2022 · paper

Two columns. Standard prompting gives a wrong answer with a red cross. Chain-of-thought prompting, with highlighted reasoning, gives the right answer with a tick.

problem → idea → payoff · contrast pair · page 1, full width

Its caption opens: “Chain-of-thought prompting enables large language models to tackle complex arithmetic, commonsense, and symbolic reasoning tasks.”

The two columns are identical except for the highlighted text, so the eye lands on the idea. ✗ becomes ✓.

Borrow: Hold everything fixed except your idea, and colour only that.

Figure 1, Wei et al. CC BY 4.0.

Lead with the idea

Use this when one idea carries the paper.

CLIP · ICML 2021 · paper

Three numbered panels: contrastive pre-training on a grid of images against captions, building a classifier from class names, and zero-shot prediction on a photo of a dog.

idea → payoff · strip · page 2, full width

Its caption opens: “Summary of our approach.”

Three numbered panels show how the model is trained, and then how it is used.

Borrow: Number the panels, and end on the use.

Figure 1, Radford et al. From the authors’ repository, MIT licence.

MAE · CVPR 2022 · paper

A flamingo photo with most patches masked; the visible patches enter a large encoder; a small decoder rebuilds the full image.

idea → payoff · strip · page 1, one column

Its caption opens: “Our MAE architecture.”

An architecture with five elements and one real image. The big encoder takes few patches; the small decoder takes all.

Borrow: Draw only what is new, and let size mean something.

Figure 1, He et al. CC BY 4.0.

LIME · KDD 2016 · paper

A model predicts flu from a patient's symptoms; a magnifying glass labelled LIME highlights sneeze and headache; a doctor makes the decision.

problem → idea → payoff · strip · page 2, full width

Its caption opens: “Explaining individual predictions.”

One patient tells the whole method, and the figure ends at the doctor who decides.

Borrow: End at the person who uses the output. Watch out: Generic icons. Use a real case today.

Figure 1, Ribeiro et al. © the authors.

Generative Image Dynamics · CVPR 2024 · paper

A photo of a rose, then colourful motion maps at several frequencies, then a looping video and interactive dynamics shown as time slices.

idea → payoff · strip · page 1, full width

Its caption opens: “We model a generative image-space prior on scene motion: from a single RGB image, our method generates a spectral volume …”

The middle panel shows what the new representation looks like, where most papers put a labelled box.

Borrow: Replace boxes with pictures of what is inside them.

Figure 1, Li et al. © the authors.

AnySat · CVPR 2025 · paper

Five satellite datasets with different sensors and resolutions flow into one box labelled AnySat, which feeds five tasks with real output maps.

problem → idea → payoff · funnel · page 1, one column

Its caption opens: “Multi-Dataset Training.”

Many real inputs, one model, many real outputs. The shape is the claim.

Borrow: For generality, draw the funnel and fill both ends with real samples.

Figure 1, Astruc et al. CC BY 4.0.

Swin Transformer · ICCV 2021 · paper

Two stacks of feature maps. Swin has three scales and serves classification, detection and segmentation. ViT has one scale and serves classification only.

idea → problem · contrast pair · page 1, one column

Its caption opens: “(a) The proposed Swin Transformer builds hierarchical feature maps by merging image patches (shown in gray) in deeper layers …”

Ours and the nearest baseline are drawn with the same shapes, so the only difference is the contribution.

Borrow: Draw the baseline the same way as your method.

Figure 1, Liu et al. CC BY 4.0.

YOLO · CVPR 2016 · paper

A photo of a person, a dog and a horse; one network; the same photo with a labelled box on each. Three steps are listed.

idea → payoff · strip · page 1, one column

Its caption opens: “The YOLO Detection System.”

The claim is simplicity, and the figure is one photo, one network and three steps.

Borrow: If simplicity is the point, stop at three parts. Watch out: Tiny in print.

Figure 1, Redmon et al. © the authors.

LLMTime · NeurIPS 2023 · paper

A language model completes a sentence and, below it, a string of numbers that is a forecast. Two small plots on the right show results.

idea → payoff · strip · page 2, full width

Its caption opens: “We propose LLMTime, a method for time series forecasting with large language models (LLMs) by encoding numbers as text and sampling possible extrapolations as text completions.”

Text above, numbers below, the same model between. The analogy is the idea, and two small plots pay it off.

Borrow: For “treat X as Y”, draw X and Y in parallel.

Figure 1, Gruver et al. © the authors.

Lead with the payoff

Use this when the result speaks for itself: a new ability, or a clear win.

3D Gaussian Splatting · SIGGRAPH 2023 · paper

Six crops of one bicycle wheel: three earlier methods, the paper's method twice, and ground truth. Each carries its frame rate, training time and PSNR.

problem → payoff · comparison row · page 1, full width

Its caption opens: “Our method achieves real-time rendering of radiance fields with quality that equals the previous method with the best quality …”

One scene for every method, with speed, training time and quality printed on each image.

Borrow: Put the numbers on the picture, and include the ground truth.

Figure 1, Kerbl et al. CC BY 4.0.

EfficientNet · ICML 2019 · paper

Accuracy against number of parameters. A red line of EfficientNet models rises above dashed lines of other named models.

payoff · plot · page 1, one column

Its caption opens: “Model Size vs. ImageNet Accuracy.”

The claim is a trade-off, so the figure is the trade-off: one red line above every named rival.

Borrow: Plot accuracy against cost, and label the points directly.

Figure 1, Tan and Le. From the authors’ repository, Apache 2.0 licence.

CycleGAN · ICCV 2017 · paper

Image pairs translated both ways: Monet and photos, zebras and horses, summer and winter. Below, one photo in four painters' styles.

payoff · gallery · page 1, full width

Its caption opens: “Given any two unordered image collections X and Y, our algorithm learns to automatically “translate” an image from one into the other and vice versa …”

Examples anyone can judge in a second, and a two-way arrow for the method’s key property.

Borrow: Choose examples a stranger can judge.

Figure 1, Zhu et al. © the authors.

pix2pix · CVPR 2017 · paper

Six input and output pairs: labels to street scene, labels to facade, black-and-white to colour, aerial to map, day to night, edges to photo.

payoff · gallery · page 1, full width

Its caption opens: “Many problems in image processing, graphics, and vision involve translating an input image into a corresponding output image.”

Six problems solved the same way, each as an input and an output under a two-word title.

Borrow: For breadth, repeat one small unit.

Figure 1, Isola et al. © the authors.

DINO · ICCV 2021 · paper

Eight photos, each beside an attention map in which the main object glows.

payoff · gallery · page 1, full width

Its caption opens: “Self-attention from a Vision Transformer with 8 × 8 patches trained with no supervision.”

The finding is a surprise, so the figure shows only the surprise: objects light up without any labels.

Borrow: Show the unexpected behaviour on many inputs.

Figure 1, Caron et al. CC BY 4.0.

ImageBind · CVPR 2023 · paper

Three numbered boxes: a sound retrieves images, depth and text; a bird photo plus wave sounds gives birds in water; sounds generate images.

payoff · gallery · page 1, full width

Its caption opens: “ImageBind’s joint embedding space enables novel multimodal capabilities.”

Three new abilities in three numbered boxes, shown before the model is explained.

Borrow: Give each new ability a box and a three-word title.

Figure 1, Girdhar et al. © the authors.

GeoChat · CVPR 2024 · paper

One aerial image in the centre with six dialogue boxes around it, each a different task.

payoff · hub · page 1, one column

Its caption opens: “GeoChat can accomplish multiple tasks for remote-sensing (RS) image comprehension in a unified framework.”

One aerial image with six tasks around it. Colours tie each answer to its region.

Borrow: Hold the input fixed, and vary the question. Watch out: Dense at column width.

Figure 1, Kuckreja et al. CC BY 4.0.

Lead with the scope

Use this for datasets, benchmarks and papers with several contributions.

Segment Anything · ICCV 2023 · paper

Three panels labelled task, model and data: promptable segmentation, the model's three parts, and a data loop giving one billion masks.

scope · triptych · page 1, full width

Its caption opens: “We aim to build a foundation model for segmentation by introducing three interconnected components …”

Three contributions, three panels, and the same three names as the introduction.

Borrow: Give each contribution a panel, with matching words.

Figure 1, Kirillov et al. © the authors.

SatlasPretrain · ICCV 2023 · paper

A globe with arrows to real satellite tiles for deforestation, wind turbines, floods, vessels and more, and a box of key statistics.

scope → payoff · hub · page 1, full width

Its caption opens: “SatlasPretrain is a large-scale remote sensing dataset.”

A globe for coverage, real tiles named by their use, and the counts in one box.

Borrow: For a dataset: real samples, real uses, three or four counts.

Figure 1, Bastani et al. CC BY 4.0.

SWE-bench · ICLR 2024 · paper

A GitHub issue and codebase, a language model that writes a pull request, and unit tests that fail before and pass after.

scope → payoff · strip · page 1, full width

Its caption opens: “SWE-bench sources task instances from real-world Python repositories by connecting GitHub issues to merged pull request solutions that resolve related tests.”

It follows one real issue from input to score, so the task and the metric are learnt together.

Borrow: For a benchmark, show one item end to end.

Figure 1, Jimenez et al. CC BY 4.0.

SustainBench · NeurIPS 2021 · paper

Coloured blocks, one per Sustainable Development Goal, each listing tasks with an input image, an output image and a count of examples.

scope · gallery · page 2, one column

Its caption opens: “Datasets and tasks included in SustainBench ranging from poverty prediction to land cover classification …”

Fifteen tasks, grouped and coloured by the goal each serves, each with its input, output and size.

Borrow: Group many items by something the reader cares about. Watch out: Near the limit for one figure.

Figure 1, Yeh et al. © the authors.

CheXpert · AAAI 2019 · paper

Two chest X-rays feed a model, which gives a probability for each of 14 findings.

scope → payoff · strip · page 1, one column

Its caption opens: “The CheXpert task is to predict the probability of different observations from multi-view chest radiographs.”

One real case defines the task: these X-rays in, these fourteen probabilities out.

Borrow: Show one real input and its full output.

Figure 1, Irvin et al. CC BY 4.0.

What the 24 have in common

  • Page 1. 20 of 24 put Figure 1 on the first page. The other 4 put it on the second.
  • Real data. 20 of 24 are built from real images, series or text, not icons.
  • Few parts. Most have two to four groups. A gallery repeats one small unit.
  • Rarely an architecture. Only MAE calls itself one, and it has five elements.
  • Captions differ. 7 of 24 open with a label, such as “Our MAE architecture.” The strongest open with the claim itself, as Chain-of-Thought, DINO and ImageBind do.
  • Few numbers. Only 5 carry a measured result. The rest show outputs that anyone can judge by eye. A forecast or a temperature cannot be judged by eye, so our figures need the number.
Paper Leads with Layout Page Real data Number Strongest on
Informer problem plot 1 yes result problem framing, evidence
ResNet problem plot 1 yes result clarity, problem framing, readable in print
Chain-of-Thought problem contrast pair 1 yes none clarity, novelty, real examples
CLIP idea strip 2 partly none novelty, hierarchy, stands alone
MAE idea strip 1 yes none novelty, readable in print, clarity
LIME idea strip 2 no none clarity, hierarchy
Generative Image Dynamics idea strip 1 yes none hierarchy, novelty
AnySat idea funnel 1 yes none real examples, clarity
Swin Transformer idea contrast pair 1 no none novelty, problem framing, readable in print
YOLO idea strip 1 yes none clarity, novelty
LLMTime idea strip 2 yes result novelty, evidence, real examples
3D Gaussian Splatting payoff comparison row 1 yes result evidence, stands alone, problem framing
EfficientNet payoff plot 1 yes result evidence, clarity, readable in print
CycleGAN payoff gallery 1 yes none real examples, clarity
pix2pix payoff gallery 1 yes none real examples, hierarchy, stands alone
DINO payoff gallery 1 yes none stands alone, real examples
ImageBind payoff gallery 1 yes none real examples, hierarchy
GeoChat payoff hub 1 yes none real examples, stands alone
Segment Anything scope triptych 1 partly counts hierarchy, stands alone
SatlasPretrain scope hub 1 yes counts real examples, evidence
SWE-bench scope strip 1 yes none real examples, clarity
SustainBench scope gallery 2 yes counts hierarchy, real examples
CheXpert scope strip 1 yes none clarity, real examples
BLINK scope hub 1 yes none real examples, hierarchy

Number: the figure carries a measured result, only counts of data, or no number.

Make your own

  1. Write the sentence. “Today, [problem]. We [idea]. Now, [payoff with a number].”
  2. Choose what leads, from the table above.
  3. Follow one real example through every panel. Pick a typical case, not the best of fifty.
  4. Sketch on paper first, in week one, with three or four groups and one reading direction.
  5. Replace boxes with pictures. A box that says “Encoder” shows nothing.
  6. Put one number on it, with the baseline beside it.
  7. Colour only what is new. Everything else is grey.
  8. Build it at final width, as a vector PDF, with text of at least 6 to 7 pt. Put it on page 1.
  9. Open the caption with the claim, and use the abstract’s words on the figure.
  10. Test it for ten seconds on someone outside the project. Redraw what they miss.

Writing Papers works through these steps for five papers.

Checklist

The figures are reproduced from the cited papers for teaching and criticism only. Each belongs to its authors and publisher. If you are an author and would like a figure removed, please write to us.