I wrote a previous post explaining how I tried to use AI responsibly when creating scientific figures. Since I wrote that, I have finished my first major project using this method (an ASABE conference paper). As so often happens, I discovered a variety of complications, pain points, and other minor issues with my original workflow, to the point where I think a follow-up is warranted.

For those of you who didn’t read the first post (and are too lazy to now), here’s the TL;DR:

• Generate image assets for the figure (icons, etc.) using an image generator
• Clean up the results manually
• Composite in Inkscape with appropriate text

This hasn’t really changed, but I’ve found a couple tricks to improve the speed and efficacy of this process which are potentially worth sharing.

Image Generation

I was previously using Google’s Nano Banana model that’s built into Gemini. This is a very good image generator, but it’s a little frustrating to use for this application. One obvious problem with Gemini is that I don’t want to pay for it, and I was worried about reaching the usage limits of the free tier. Therefore, I decided to give OpenAI’s image generator a shot, seeing as Dr. Li could pay for it instead: we have an OpenAI platform account for the lab.

I quickly found that OpenAI’s image generation playground is extremely suitable for my needs. This is not necessarily because the generated results are better. It’s more that it offers far more control over the results compared to Gemini. For instance, one of my favorite features is that I can force it to generate images with a transparent background:

This saves a common and annoying post-processing step that I previously had to do manually. As it happens, there are some other convenient settings that I can change, including the ability to set the output aspect ratio (sometimes you want landscape instead of square), and an automatic setting to make it generate multiple images at once. I usually do this and pick whichever one I like best.

Reference Images

One of the most useful techniques that I discovered was the ability to give the model reference images to work off of. For instance, at one point, I found myself with a need to make two different icons to represent annotated and unannotated data in a flowchart. Logically, I wanted these icons to feature very similar designs. However, naively trying to generate both of them separately led to totally different results.

There’s a feature in the image generation playground were you can upload a reference image to include with your prompt. So I tried first generating one icon, and then uploading that icon as reference when generating the other one. Here was the prompt that I used:

Here are the side-by-side results. I was quite pleased with these!

This can also be a useful method when trying to create an icon base on a real-life item. For instance, we have a robot that has a distinctive design, and I wanted a custom icon for it that I could use in my figures. To accomplish this, I provided a CAD model of the robot as reference.

It turns out you can also use this technique for general-purpose image editing. Dr. Li handed me a PNG image of a figure that contained text, which he wanted to be able to edit:

Not wanting to spend hours in Affinity Photo carefully removing and re-adding the text, I simply added a bunch of black boxes on top like it was an Epstein email that mentioned Trump:

After that, I went into the image generation playground, and asked it politely to remove all the black boxes:

I was worried this might just add new gibberish text, but it didn’t!

After this, of course, it was a trivial matter to copy this into PowerPoint and add text boxes in all the appropriate places. Dr. Li never said anything about it, which, knowing him, means that I did a good job.

Vectorization

I know that I initially wrote off Inkscape’s built-in tracing tool as not very useful. It’s true that if you want to convert a very complex raster to vectors, it’s probably not a good bet. However, I did figure out one situation in which it works perfectly: If you deliberately take a colored icon and trace it in “single scan” mode, you usually get a very good result which you can then go back and add color to manually.

I found later that it’s even possible to get decent results for colored icons, as long as the colors are in uniform blocks. That said, you might have to play with the sliders a bit, especially for more complex input. The main thing to note is that I always check the “stack” box. Doing this sometimes leads to weird fringing effects, but I find these less distracting than the gaps that you get between colored areas when you leave it unchecked.

Conclusion

Using these techniques, I’ve found that I’ve been able to produce much higher quality figures than before. The downside, ironically, is that, if anything, I’ve been spending more time making these figures than I have in the past. If AI is supposed to make people more productive, that has definitely not been my experience. That being said, it would be more time consuming still to produce figures of this quality without AI.

It’s important to emphasize, though, that it’s still early days for this. I’m still working through this process and finding the tools and workflows that I’m most comfortable with. What works for me might not work for everyone. For example, I’m already very comfortable using Inkscape and I don’t mind putting a lot of good old-fashioned non-AI-assisted graphic design work into my figures. Obviously, not everyone will be willing/able to do this.

All that being said, I’m definitely going to keep iterating on this technique. I think that it’s working! Although I guess the real test of that will be whether my ASABE paper wins an ITSC award this year.

Just for fun, here’s a little gallery of the figures I’ve made in the past month: