How to create a graphical abstract using AI (Or not?)

Chanelle Lai
58 minutes ago
7 min read

How many times have you tried submitting a paper to a high profile journal, only to realise that you’re required to include a…

Wait for it…

“Graphical Abstract”?

Seth Meyers sitting at a desk, looking overwhelmed and counting rapidly on his fingers as if listing a never-ending series of items.

Too many to count, right?!

Now you must be thinking…

“I’m a scientist, not a graphic designer!! “

So you try making one from scratch, but sometimes inspiration fails to strike…

A close-up meme of a white cat looking blankly into the camera with the text "NO THOUGHTS TODAY, HEAD EMPTY."

Sometimes starting is the hardest step!

You try searching up template designs, but you have already seen three different graphical abstracts with the same design…

A minimalist graphical abstract template in shades of grey and blue. It features a three-column layout with headers for "Background," "Methods," and "Results," along with placeholders for a main title, DOI, article title, and authors.

I'm bored already!

So you turn to the next thing, getting AI to make a graphical abstract for you (or if you’re a big fan of using AI for things, maybe this was step 1).

But does using AI to make a graphical abstract actually work? Does it hold up against the expectations for a typical graphical abstract? Can it produce something purposeful, accurate, journal-accepted and, well... actually good?

The other question to consider is of course which AI tools are best to use for crafting graphical abstracts? (So many to choose from!)

Well don’t worry, because we have tested them FOR YOU!

We’ve used the free versions of the AI tools mentioned below for this little experiment, because not everyone wants (or needs!) a paid subscription for these tools.

So it begins…

Let’s use the paper below as an example; something new and niche, just like what you’re working on!

Title: Cell cycle checkpoint activity in the malaria parasite Plasmodium falciparum

Here is a link to the full paper if you are interested.

I picked this preprint because I (Chanelle) had just read it for a journal club discussion.

Time to make a graphical abstract!

An animated GIF of a scientist in a white lab coat and blue gloves picking up safety goggles from a lab bench. Bold blue text at the bottom reads "LET'S DO THIS."

Step 1: NotebookLM

Here is where we introduce our first tool, NotebookLM by Google. It is designed specifically to act as your personalised research assistant. It excels at distilling key messaging from a dizzying amount of multiple sources (and can even create slide decks, infographics or podcasts from your sources!). What sets it apart from Large Language Models like ChatGPT, Claude or Gemini is that it only reads the documents you give it. So, it heavily reduces the risk for AI hallucinations. Since it can create infographics, it is a great first stop for getting a sense of what your own graphical abstract can look like (or shouldn't look like!).

Let’s add in our source material (your full paper in PDF) and see how it goes…

An animated GIF of the characters from Little Einsteins in their rocket ship, enthusiastically waving their arms and preparing for a journey.

Here we go!!

BUT WAIT! Just adding one source and hitting GO might not work so well. For instance, this is what it ended up generating for our example paper:

An AI-generated infographic from NotebookLM titled "Unlocking Malaria’s Hidden Checkpoints." The design is split into a green "Discovery" section and a dark blue "Impact" section. It features stylized illustrations of malaria parasites, DNA strands, and histone markers, along with three speedometer-style gauges at the bottom comparing drug sensitivity across different parasite strains.

It doesn't look half bad and is a good conceptual start, but on closer inspection it has plenty of irrelevant information and a ton of crazy graphics that are a far-cry from scientifically accurate and straight up WRONG.

How can we fix this mess?!?

Let us try adding some photos as reference for what the parasite should like, and then telling the tool what the summary (and key takeaway) of our paper is!

IMPORTANTLY: If you have read any of our previous blog posts on visual design, you would know that we value negative space on our graphical abstracts. That is why one of the subsequent prompts to the AI was to include 20-30% of negative space, which helps prevent the usual AI clutter.

Click here if you would like to read our previous post on creating a graphical abstract!

After adding some much-need context to our prompt, this is what we ended up with…

An AI-generated scientific infographic from NotebookLM with several red arrows and text annotations pointing out scientific errors. The annotations critique an inaccurate title, anatomical mistakes like the parasite having too many rhoptries, nonsensical bar graphs, and misleading illustrations of inhibitor effects. The image serves as a "critique layer" to demonstrate why AI outputs require expert fact-checking.

Slightly better, but there are still a LOT of critical information that it’s not getting right… 😞

Here is a list of subsequent instructions that we used to further tweak the prompt, including design-related guidance to steer the tool in the direction we want.Prompt list:

Don’t use tables in the infographic.
Use a uniform colour scheme with only 3 colours.
Make sure that all scientific names are properly italicised.
Make the text as concise as it can be.

This was the result:

An AI-generated infographic from NotebookLM titled "Cell cycle checkpoint activity in the malaria parasite P. falciparum." The layout is organized into three vertical panels on a light cream background: "The Cryptic Checkpoint Mechanism," "Comparative Potency of PIKK Inhibitors," and "Strain Sensitivity and Resistance." It includes illustrations of parasites, protein structures, a DNA helix, and horizontal bar charts showing IC50 variations across different malaria strains.

Still far from perfect, but after the added instructions, it has certainly increased in visual appeal, focused clarity and layout simplicity. Accuracy of information still remains a big issue however…

Because Google’s Gemini is such a powerful image generator, we wondered whether using it, rather than NotebookLM alone, would make any difference to the results.

Step 2:Google Gemini image generation

We started with the written summary from NotebookLM and added plenty of extra context. Once we realised these AI tools still need a lot of guidance, we wrote a much more detailed prompt.

The instructions included:

The paper's summary.
Use 20-30% of negative space in the infographic.
Use more minimalistic and realistic art.
Infographic in landscape orientation.
Use title from the paper for the infographic.
No tables in the infographic.
Use a maximum of two to three colours for the infographic’s colour scheme.
Make sure that all scientific names are properly italicised.
Make the text as concise as it can be.
Ensure that the title of each subsection is intuitive to the reader.

And here were the results…

A detailed AI-generated infographic titled "Unveiling the Cryptic Checkpoint: Plasmodium PIKK Homologues and Inhibitors." The central feature is a large diagram of a malaria parasite cell with labeled organelles like the nucleus, rhoptries, and apicoplast. Surrounding panels describe the Plasmodium life cycle (Ring to Merozoite), experimental inhibitors KU-55933 and VE-821 with their chemical structures, and the impact of these inhibitors on schizogony and merozoite formation.

WOAH!! Talk about information overload!!! 😟

A grainy, handheld-style GIF of a small, fluffy grey tabby kitten. The kitten looks up with large, dark eyes and moves its head slightly, appearing curious and playful.

POV: You trying to read this graphical abstract

So as a quick-fix one-shot solution? Not the best tool (in our humble opinion), for generating graphics from new information.

After several attempts of trying to get Gemini to declutter the graphical abstract we ended up with this:

An AI-generated infographic titled "Cell cycle checkpoint activity in the malaria parasite P. falciparum." The graphic is divided into three sections: "Parasite DNA Monitoring," which highlights the absence of traditional kinases; "Challenging Parasite Kinases," showing the potency of inhibitors VE-821 and KU-55933; and "Altered Drug Sensitivity," which explains strain variations and the impact of the Kelch13 mutation on drug resistance.

Still too much text, too much clutter, and too little rest-for-the-eyes space…

At this point, you might be thinking, “You’re just chucking full papers at random AI tools to generate these infographics. Of course they won’t be good”, or “ChatGPT can do so much better!”. But be patient. We’re on a journey.

Now, let’s see how some of the other biggest AI tools, powered by different underlying LLMs, compare!

An animated GIF of a man (Mike Hitt) looking upwards with an exhausted expression, rubbing his face with both hands in a gesture of deep frustration. Bold black text to the right reads "HERE WE GO AGAIN."

Step 3: Comparison between Gemini, (ChatGPT and Microsoft Copilot)

For this comparison, we took no chances and created a much longer, more detailed prompt, which included….

Landscape layout
Negative space minimum 20-30%
Colour scheme
1. #0090A0 (Dark Teal)
2. #1C9BB3 (Medium Teal)
3. #B9E2ED (Light Blue)
4. #FFFFFF (White)
5. #000000 (Black)
6. #FABE00 (Accent: Bright Yellow)
Font
1. Montserrat Bold/ Regular
Image provided as an example of layout for the graphical abstract (see below)
Never include more than 80 words in the graphical abstract.
Make sure all scientific names are properly italicised.
Use minimalistic and realistic art.

Alt text: A colorful teal and yellow scientific poster template featuring a large illustration of a cyclist in the bottom left. The layout includes placeholders for a background, key findings with a line graph, methods, and a takeaway section, all written with humorous filler text such as "Dr. Smart, Dr. Handy, Prof. Had-nothing-to-do-with-it." — Sample layout offered as reference

Gemini

To fairly compare the best that Gemini has to offer with its rivals, we wanted to use Gemini's customisable project-function (called "Gems"), where you can create a tailored 'Gem' with specific instructions Now, an important thing to note is that in Gemini, you can create what is essentially a canned template for your work, called a “ Gem”.

So instead of having to send the same prompt repeatedly for each new project, you could have something like this:

A screenshot of the Google Gemini "Gems" interface titled "AYS Graphical abstract." The interface shows a list of custom instructions for the AI, such as using 30% negative space, italicizing scientific names, and keeping word counts under 90. Two image templates are uploaded in the "Knowledge" section to provide a visual reference for the desired layout.

Creating a “Gem” with embedded instructions for your graphical abstracts.

Using the Gem we made, the graphical abstract generated looked something like this:

Looks pretty much like a standard graphical abstract, but that is still not what we are looking for. We want excellence!

Because we wanted the graphical abstract to feel closer to our templates in terms of layout, we asked it to rearrange the information to more closely match one of the reference graphic provided in the Gem and…

An AI-generated graphical abstract fails by hallucination, featuring a large, irrelevant illustration of a cyclist overlaid on top of scientific data. The graphic includes a teal header titled "Understanding the Plasmodium falciparum DNA replication checkpoint and Kelch13 mutation," with flanking panels of bar graphs and genetic diagrams that are partially obscured by the cyclist. A humorous quote reads: "Scientific posters don't have to have to ugly."

Why is there a cyclist?!?

An animated GIF of Cynthia Erivo on a red carpet, looking intensely shocked and wide-eyed with her mouth slightly open, capturing a moment of pure disbelief.

And now here is another important lesson…

Be very VERY specific with your prompts, because when we asked Gemini to replace the cyclist with a photo of the parasite, this happened:

An AI-generated graphical abstract demonstrating a persistent "cyclist hallucination." Despite adding a parasite image, the AI kept the large illustration of a man on a bicycle, which partially overlaps with scientific data on Plasmodium falciparum DNA replication and drug resistance graphs.

And when we asked Gemini to remove the cyclist entirely:

An AI-generated graphical abstract showing a failed attempt to remove a hallucination. While the rider has been removed, a large grey silhouette of a bicycle remains in the bottom left corner, floating next to an actual parasite micrograph and obscuring the intended scientific layout.

Well, can’t say it didn’t do what we asked it to.

An animated GIF of the Muppet Elmo looking directly into the camera with a wide-eyed, startled, and frozen expression, as if he has just seen something shocking or confusing.

A lot of back and forth later and we were eventually hit with an image creation limit, leaving us with this:

An AI-generated graphical abstract where the cyclist hallucination has finally been replaced with a real micrograph of a Plasmodium parasite. The layout is cleaner, featuring panels on checkpoint hallmarks, inhibition mechanisms, and Kelch13 mutation drug resistance, though some minor AI-generated text artifacts remain.

Not the worst one, but definitely nowhere near our expectations.

Chat GPT

One of the more commercialised AI LLMs that has changed the way we view AI forever, this is the one AI tool most people would expect to do well. You’ll see pretty soon, though, that this is not the case at all.

Using the same prompts as for Gemini, this is what it generated:

A sophisticated AI-generated graphical abstract with a central, 3D textured illustration of a Plasmodium schizont. The surrounding panels use a clean blue and tan color palette to describe the absence of ATM/ATR kinases, the presence of PI3K/PI4K, checkpoint hallmarks (BrdU and H2A-P signals in petri dishes), inhibitor actions of KU-55933 and VE-821, and a bar chart at the bottom comparing resistance factors across different parasite lines.

…Yeah.

Hardly believable, plus this is the kind of image you would imagine when you think “AI generated”.

We gave it the same cyclist poster as a design template, and asked for it to be more minimalistic and scientifically acceptable. We also asked it to remove the petri dish image.

This is what it ended up with:

A teal-themed AI-generated graphical abstract titled "Checkpoint-Independent DNA Replication Mechanisms in Plasmodium." It features a central schizont illustration with dashed lines connecting to four sections: genomic checkpoints (ATM/ATR), inhibitor actions of KU-55933 and VE-821, hallmark graphs of DNA replication, and resistance factor bar charts illustrating the impact of the Kelch13 mutation.

Chat GPT has in fact, removed the petri dish. Also, apparently now it is unable to spell…

Can’t believe I’m saying this, but maybe the image creation should be left to the scientist… (Or the scientific illustrator! 😉)

A colorful, hand-drawn animation of a teal bird wearing a crown and a purple robe, diligently painting on a canvas with a paintbrush.

And finally…

Microsoft Copilot

Most people have it, nobody uses it. But maybe you should? 😃

We gave it the same prompts from before, and this is what it came up with:

A high-detail, medical-style illustration of a P. falciparum schizont surrounded by four informational panels. The panels use a warm tan and blue color scheme, featuring chemical structures for inhibitors KU-55933 and VE-821, a DNA double helix, and a bar graph showing relative IG50 levels.

Probably the cleanest design we’ve got back in the first go AND hey, it can spell!

The image certainly leaves much to be desired, but let’s see if it can be fixed with a few more instructional prompts…

A clean, light-blue infographic titled "Checkpoint-Independent Replication Arrest in Plasmodium falciparum." It features a central 3D diagram of a schizont, flanking panels comparing human and Plasmodium genomes, chemical structures of inhibitors, and a bar graph showing resistance factors in different parasite strains.

Still not great, but less bad!

In summary:

So, what did we actually learn from this whole adventure?

1. Start with a strong concept

Use NotebookLM to distil the key message from your paper before touching any image tool. Going in without a clear visual brief is like asking someone to draw your research from a vibe. Spoiler: it doesn't go well.

2. Be obsessively specific with your prompt

The cyclist did not appear by accident. If you don't spell out every detail, the AI will fill in the gaps creatively. Very, very creatively. List your layout, colours, word count, and scientific naming conventions upfront, and be prepared to iterate.

3. Plan to do some finishing yourself

AI outputs are static images, so every typo, wrong label, or rogue cyclist means starting over. Use AI to get most of the way there, then finish the job yourself in a tool like Affinity Designer.

Main takeaway

Think of AI as a very enthusiastic intern: great at reading, summarising, and nodding along, but not quite ready to be left alone with Illustrator. And if we're being honest, in the time we spent feeding prompts into these tools, we could have finished the graphical abstract ourselves...

No one knows your research better than you do. So experiment with AI, use only what works, and tell your story your way.

If you would like to learn more tips on how to create an EXCELLENT graphical abstract, here are some more light reading you can do:

How to design an effective graphical abstract: the ultimate guide

Best examples of graphical abstracts

Want to learn more about using AI in your research?

If you'd like to go deeper into how AI can support your work as a researcher, ethically, effectively, and without the headaches, we have just the right resource for you.

Check out our course AI for Researchers: Unlocking New Potentials and Avoiding Pitfalls

About the author

Hello there! My name is Chanelle, and I am a PhD candidate whose research focuses on characterising proteins on the surface of Plasmodium parasites, which can cause malaria. I have a keen interest in all things science communication, and this is my first blog post that I have written as part of my internship here at Animate Your Science! Hope you’ve enjoyed the read!