--- title: Tips For Using AI-Generated Images In Your Slides teaser: I used AI to generate images for my slides, but it required some work to get them to look how I wanted. Here's what I learned along the way. tags: artificial intelligence,images,slide deck author: Matheus Richard published_on: 2024-05-08 --- I gave a talk on [Tropical.rb] and, while writing it, I had a goal of [not using bullet points] in my slides. That made me create *a lot* more slides, and many of them contained images. I used AI to generate the illustrations for my presentation, but it required some work to get them to look how I wanted. Here's what I learned along the way. ## Getting What You Want The first step when generating images is figuring out what you *actually* want. We might get stuck on analysis paralysis, so try to get something on the screen as soon as possible. ### Iterate Over Ideas I think that was one of the best parts of using AI. Sometimes, I had to illustrate some abstract concepts like "HTTP requests", so I didn't know what image I wanted in the first place. With AI, I could just give it a basic prompt like A drawing representing an HTTP request, and it would at least get something on the screen. Once I see it, I can get inspired and iterate on that idea. Play around with prompts a bit and iterate over the results. You *probably* won't get it right the first time.

Four pictures of generated images. The first one has a monk with red skin. The second one has a monk with a ruby in his forehead. The third one as an anthropomorphic ruby gem with beard and one hand raised. The fourth one is an anthropomorphic ruby gem with a beard, holding a stick on one hand a with a finger raised on the other. — Sometimes you evolve your ideas. Sometimes you evolve your prompts.

Some things were very easy to generate, while others were impossible. For example, I wanted a cartoon version of ["The Thinker" statue], so I just prompted The Thinker statue, cartoon style and got exactly what I wanted on the first try.

On the other hand, I wanted to generate a version of [the "Drake meme"], but because it references a real person, the AIs wouldn't generate it. It was so frustrating because I didn't even want Drake himself in the image, just the meme format. I ended up giving up on this one. ### Be specific The more specific you can get, the easier will be to get what you want. I'm not an artist, so sometimes was hard to convey what I wanted because I didn't know words for it ([LLMs][] like ChatGPT to help you figure that out). Try using references like "anime style", "Van Gogh", "synth-wave", "cyberpunk", etc. You can be even more specific specifying a particular piece of art or sub-genre (e.g., chibi manga). Using a particular artist's name can also help, but the AI might refuse to generate it for copyright reasons.

A dog wearing a ninja outfit in cartoon style — Prompt: "A ninja dog, Masashi Kishimoto style, low detail"

Negative prompts are very helpful too. If the AI is doing something you don't want, tell it to stop. For example, append no text, no background, or don't draw faces to your prompts. ### Dealing With Text While it's much better than it once was, AI still often struggles with text. For instance, it wouldn't generate an image with the word "async" correctly spelled. It would either skip the N or merge the Y and N in a weird way that looked like a new letter. I tried prompting "ASYNC" in uppercase, but it **did not** help at all.

A cartoon of a kid looking at a big rock with the word ASYC written in it — Prompt: "A small kid looking at a big rock with the word ASYNC written, no background, cartoon style, simple"

I told it to be super careful with the letters, and that improved the results a bit, but it made the AI add that phrase to the image.

A cartoon of a kid looking at a big rock with the word ASYNC written in it, with a conversation balloon that reads 'Be super careful' — Prompt: "A small kid looking at a big rock with the word ASYNC written, be super careful with the letters, no background, cartoon style, simple"

The solution for me was to prompt do not misspell.

A cartoon of a kid looking at a big rock with the word ASYNC written in it — Prompt: "A small kid looking at a big rock with the word ASYNC written, do not misspell, no background, cartoon style, simple"

## Dealing With Small Images I wanted images to use in my slides, so they either had to be big and in a 16:9 aspect ratio, or have no background so I could use them with a flat color background. Most of the AIs available, in particular the free ones, only generate small images, or they only generate square images. If that doesn't work for you, here are some ways to work around that. If you know how to use image editors, that will be useful, but you don't need to be an expert to do many of these things. ### Cropping It might be possible to just display a part of the image, or the image in a smaller size. If that's the case, you can just crop it. In my case, I want big images in my slides, as they are projected on a big screen and there was nothing else in the slides. I wanted the images to blend with the background, so I didn't want a hard cut on the edges. One of the simplest tricks you can use is to crop the image in a circle or rounded rectangle. Something about cropping images in circles makes the cut look intentional. Check this slide example:

a grassy landscape with ruby gems scattered around like flowers and a sunsetting sky — Prompt: "a landscape with ruby gems, cartoon style, no background, simple"

If you have some skills with image editors, you can also make some parts of the image go outside the circle, which has a nice effect.

a cartoon picture of a car track with high traffic, with a red bullet train runs on a track on top of it — Prompt: "a car track with high traffic, but a red bullet train runs freely on the top, no background, cartoon style, simple, flat colors"

### Removing Background Images with transparent backgrounds are really useful on slides because you can use them on top of a flat color and they will blend nicely. Unfortunately, all the AIs I tried cannot generate images with transparent backgrounds. If you ask for that you might get a "fake PNG" (you know, those with a checkerboard background, not actual transparency).

A cat wearing a ninja outfit in cartoon style with a checkerboard background — Prompt: "A ninja cat, cartoon style, transparent background"

I have a few options to deal with that. The first one is to ask for images with flat color backgrounds. Then, all you need to do is to put that image on top of something that has the same color as the background. _Voilà_! You can know have this image in any aspect ratio you want.

An anthropomorphic ruby gem with a beard, holding a stick on one hand a with a finger raised on the other, in a cartoon style — Prompt: "An old ruby gem with arms and legs, who is a monk and has a beard with a finger raised giving advice, cartoon style, no background, simple"

It's possible that the "flat" background in your AI-generated image will not be super consistent: instead of a single color, you might get a gradient or many shades of the same color. If you put that on top of a flat background, you might see the edges of the image. In this example, parts of image blend with the background better than the others. The lower left corner is visibly using a different shade than the background.

A chess board cartoon picture with visible square edges separating it from the flat color background — Prompt: "a chess board from the top, no background, cartoon style, simple"

To fix that, you can blur the edges of the image a bit. I used an eraser tool with a gradient or some opacity, so the edges would blend better with flat color in the background.

A chess board cartoon picture with blurred edges — It didn't need a lot of blur to make the image blend with the background

Alternatively, you can take an image that does have a background (flat or not), and use *another AI* to remove it. I used and it worked well.

A cat wearing a ninja outfit in cartoon style with a transparent background — Note that there's still a small part of the background left there, but that's easy to fix manually. More on that later.

Another tip is to ask AI to generate images that don't touch the edges of the canvas. That allows you to easily crop or extend them, if needed. You can prompt in a vignette* to get images that live inside a "little island".

A cartoon picture of a road with a sign reading 'Side Quest'. Both on top of a floating piece of land. — Prompt: "a road with a sign that reads 'side quest', do not misspell it, cartoon style, no background, simple, in a vignette"

### Upscaling If you need higher-quality images, upscaling the image might be just enough. There are plenty of AIs available to upscale images. I like using because it can upscale up to 4 times and it has an "AI-enhance" feature that makes the final result smoother and better looking than just upscaling it.

A cartoon version of the Statue of Liberty. — Prompt: "the statue of liberty, no background, just a blue sky, cartoon style, simple"

Check the upscaled version:

A cartoon version of the Statue of Liberty with sharper details. — This might be hard to see in a small device, but the details are much sharper in the upscaled version. When projecting your slides on a big screen, this will make a difference.

Feel free to use any other upscaling service you like. They're not sponsoring me in any way (I wish they were), I just like their service. ### Extending If you need images in a different aspect ratio (like 16:9), extending the image might be the solution. There are some AIs that can extend images, but in my experience, none of them were very good. The results were inconsistent, or just plain bad. You could easily spot the generated parts because they looked different and out of place. See this example of an extended image. Both the left and right sides were extended and contain inconsistent art and overall weirdness (especially the upper-right corner).

A subway station with a big digital display, with confusing lines and colors on the sides.

If you can, try extending as few parts of the image as possible. I couldn't get this to look very good, so I worked around it by blurring the generated parts. The center of the image was the focus, so I blurred the edges. You can use any kind of blur you want (gaussian, focus, motion, etc), but in this case, I used GIMP's Mean Curvature Blur. It helped smooth out some of the rough parts, and it worked with the cartoon-styled art.

A subway station with a big digital display, all the lines are softened which makes the weird lines on the sides less noticeable.

If you pay attention, the weirdness is still there, but I did some dry runs of my talk and no one noticed it. Unless someone looks very closely, they won't notice that, so don't overthink it. It's also possible to extend an image by copying its sides. This is particularly effective on symmetric images. Adding some blurring to the copied parts can help smooth out the transitions. Here, I just cloned the sides of the image until it reached the aspect ratio I wanted.

A cartoon version of the Statue of Liberty in a landscape aspect ratio.

### Cleaning Up As you saw in the previous examples, AI-generated images are not perfect. Having some skills with image editors will help you a lot. For example, I liked one of the images generated but wanted to use a different text. I used GIMP's clone tool to erase the text, so I could add my own on top of it.

Two pictures of a subway digital display of the next station. The left one has text in the display, the right one has no text.

Another use case I had was fixing some weird stuff that the AI created. I didn't like the face in the Statue of Liberty picture in this example, so I erased some bits again using the clone tool.

Two pictures of a cartoon version of the Statue of Liberty. The left one has asymmetrical eyes, the right one doesn't have eyes at all.

As I mentioned before, images with AI-removed backgrounds may still contain some background bits. This is a good chance to fix that manually with usual image editing tools. All that said, keep in mind that, as with many things, getting it right from the source is the best way to achieve good results. ## Thoughts It's funny how AI sometimes fails to do what computers do best: being consistent. You ask for a flat color bg, and it gives you one that *looks* like one color, but it's tens of slightly different shades of that color (if you color pick, you'll get different values). Or you might ask for something very specific like Don't add text or Don't draw faces and that's exactly what it does 🤦🏽‍♂️. While it wasn't effortless to get the images I wanted, the final result was worth it. It required a lot of prompting, and some manual work to get things right but I'm happy with [the final result]. It was interesting to experience that, like [LLMs] for code, generating images also required iteration and manual work. They helped me *augment* my abilities, but they didn't replace them. [Tropical.rb]: https://www.tropicalonrails.com/archive-2024 [not using bullet points]: https://bowperson.com/2018/02/reposting-6-ways-to-quit-bullet-points-with-style-by-podium-wisdom/ ["The Thinker" statue]: https://en.wikipedia.org/wiki/The_Thinker [the "Drake meme"]: https://knowyourmeme.com/memes/drakeposting [the final result]: https://docs.google.com/presentation/d/1zwy3RZhH4TOKBTEYPqtaK2Bl_d8bcIc9VhKr9t9lasw/edit?usp=sharing [LLMs]: https://thoughtbot.com/blog/tags/large-language-models