Yellow

White

plain black

Orchestra Blue

SOLOIST

Thanks to my daughter, this past week I became one of the <a href="https://xkcd.com/1053/">Lucky 10,000</a> to learn about <a href="https://knowyourmeme.com/memes/goth-ihop">Goth IHOP</a>. After a meme went viral several years ago, somehow the Internet decided that the IHOP in Frederick, MD was 'the goth one'. So much so that Apple Maps shows it as a listing in the region:

So of course that got the staff here at AI Outtakes thinking. If there's one thing that AI art should be good at, it's depicting memes in a more literal sense.

But first, let's check in on Juggalo Sam's Club.

<code class="inline-code">a photo of juggalos, face paint, shopping at Sam's Club</code>

Okay now let's find out!

Of course, since it's the inspiration for the post we need to see what Goth IHOP might actually look like:

<code class="inline-code">a photo of a goth IHOP, exterior photo, in a strip mall</code>

<code class="inline-code">a photo of an IHOP, interior photo, goth waitstaff</code>

<code class="inline-code">a photo of a stack of goth pancakes, on a table in a diner</code>

Okay we need to see the patrons of this Goth IHOP.

<code class="inline-code">a photo of a goth IHOP, goth waitstaff, several goth guests in booths</code>

Okay, we've probably done this one to death (get it? get it? taps mic Is this thing on?). Let's see how AI does with some other popular memes.

<code class="inline-code">a photo of first world problems</code>

I wonder if DALL-E will have better luck?

<code class="inline-code">can you generate an image based on the First World Problems meme,
but in the background include a collage of things that are
ironically first world problems that the woman might be 
distressed about?</code>

Let's see about another popular meme, the <a href="https://knowyourmeme.com/memes/this-is-fine">This Is Fine</a> dog:

<code class="inline-code">a photo of a yellow dog, sitting on a chair next to a table, 
wearing a small hat, inside a burning house</code>

Moving on to 'Pikachu surprised'. Since DALL-E 3 has restrictions on representing copyrighted material we've got to go back to SDXL.

<code class="inline-code">a photo of pikachu, looking surprised</code>

<code class="inline-code">a photo of kermit the frog, sipping from a clear mug of tea, teabag in the cup</code>

Time for the final meme. The one you probably thought of when you realized this edition was going to be about memes.

<a href="https://knowyourmeme.com/memes/distracted-boyfriend">Distracted boyfriend</a>.

<code class="inline-code">a photo of a boyfriend, distracted, looking the other way</code>

<code class="inline-code">a photo of a boyfriend, with his girlfriend, looking at a different woman</code>

Can you generate a version of the distracted boyfriend meme?

<code class="inline-code">okay now can you generate a version of the distracted boyfriend 
meme but make all of the people goth</code>

That's all for this week! I hope you've had as fun reading this as I've had fun making it.

<code class="inline-code">a photo of a goth IHOP, goth waitstaff waving goodbye</code>

Generative AI through the lens of Goth IHOP.

AI Outtakes, Vol. 11

In <a href="https://blog.orchestra-ai.com/ai-outtakes-vol-84-6407fa04">last week's edition</a> as we were testing whether AI image generators could make images of men without beards or a blue suit (they cannot), we briefly tested Google's latest model, <a href="https://deepmind.google/technologies/imagen-3/">Imagen 3</a>. As the "highest quality model" from one of the world's largest companies, it should have been able to handle that prompt.

Not only did it fail, the staff here at AI Outtakes noticed something unusual. All of the people were attractive. Take the first image it generated:

This brings is to the topic of this week's edition: Can Google's Imagen 3 actually generate a picture of an ugly person? Let's find out!

<code class="inline-code">a photo of an ugly person</code>

A good (???) start! Maybe we were wrong about Imagen 3.

<code class="inline-code">a photo of an uglier person</code>

<code class="inline-code">a photo of the ugliest person</code>

<code class="inline-code">a photo of the ugliest person you have ever seen</code>

Look just because someone is ugly doesn't mean they can't dress nicely!

<code class="inline-code">a photo of the ugliest person you have ever seen, wearing a tuxedo</code>

<code class="inline-code">a photo of a person so ugly they scare pets away</code>

<code class="inline-code">a photo of a person who is ugly, but their mother says they are beautiful</code>

Aww, this looks like someone from the Manning family.

<code class="inline-code">a photo of a person who is ugly, but not like, "ugly" ugly</code>

<code class="inline-code">a photo of many ugly office workers, waving goodbye</code>

Google's Imagen 3 AI refuses to create truly ugly people, turning every attempt into an attractive portrait. See the hilarious results of our experiment.

AI Outtakes, Vol 85

The staff here at AI Outtakes have come to ABSOLUTELY DESPISE the AI Guy.

Admittedly, we were sick of him all the way back in <a href="https://www.linkedin.com/pulse/ai-outtakes-vol-5-jim-wrubel/">Volume 5</a>, in September 2023. Feels like a lifetime ago. He has not become more endearing.

Just take a look at this image we recently generated for an <a href="https://orchestra-ai.com?ref=ai-outtakes-84">Orchestra</a> client:

This photo was the inspiration for this week's edition. Is it possible for AI to generate a photo of a man with neither a beard nor wearing blue?

For this edition we ran the prompt <code class="inline-code">A photo of a clean-shaven male office worker who is not wearing blue</code> against our current favorite image generator, Recraft v3. The goal is to see how many smashes of the generate button it takes to get a photo that actually fulfills this supposedly-simple prompt.

This is pretty close! There's enough scruff there to fail the 'no beard' part, and the suit is mostly grey with some blue, but this is an encouraging first attempt!

Who would have thought the beard would be easier to fulfill than the clothes color?

Whatever the cause, this prompt has not reduced the number of blue shirts but it's dramatically increased the number of red shirts depicted.

Swing and a miss. Well, this has been a giant bust so far. 

Let's swap in a different image model. Google image gen. You're a product of one of the richest companies in the world. Same prompt. Whaddaya got?

Also no, and why is his head like a balloon on his shoulders

Anyone else starting to notice some similarities in these Google-generated people?

Last-ditch effort, because it feel wrong to use the equivalent of Albuquerque's electricity budget trying to get a photo of someone without a beard or blue suit. We previously used Black Forest Labs' Flux Pro for image gen. Can it come up with a photo of a clean-shaven man not wearing blue?

New product idea, CAPTCHA, but it's just "Pick all the men who are clean-shaven and not wearing blue".

<code class="inline-code">A photo of a company gathering featuring a photo of 300 male office workers all dressed the same. They are waving goodbye.</code>

Watch AI hilariously fail at the simple task of generating an image of a clean-shaven man not wearing blue. The results will make you laugh.

AI Outtakes, Vol. 84

One of the biggest challenges and limitations with AI image generators is the "not quite" problem.&nbsp;&nbsp;

The staff here at AI Outtakes sees this daily in our <a href="https://orchestra-ai.com?ref=ai-outtakes-83">work with Orchestra</a>. Generate an image to support a blog article or social media post and it's perfect, except... One of the people depicted has six fingers, or there's a spelling issue with generated text, or the angle is wrong, or it's too dark, or too light, or not goth enough... The list goes on.

Some issues are fixable by taking the image into an editor and using the magic erase function, but for the most part the only way to "fix" the image is to mash the generate button again and hope the next version you get doesn't have the same problem, and also it doesn't have any different problems.

Ultimately what you end up with a lot of times is spamming the magic AI button until you get something that you can live with. Not very efficient!

Recently, Google introduced their <a href="https://deepmind.google/technologies/gemini/flash/">DeepMind Gemini 2.0 Flash model</a> with multimodal capability (meaning it can work with both text and images, something OpenAI has had for over a year), and reasoning. Given it's multimodal, it gives the tantalizing promise of being able to make edits to existing images by just describing the changes you want. <a href="https://knowyourmeme.com/memes/big-if-true">Big, if true</a>!

We tested OpenAI's DALL-E 3 multimodal capability <a href="https://www.linkedin.com/pulse/ai-outtakes-vol-33-jim-wrubel-zvrde">back in April 2024</a> and found it frustratingly limited. Given almost a year and the brainpower behind one of the world's biggest companies, can Gemini 2.0 Flash do better? Let's find out!

For this experiment we're using <a href="https://huggingface.co/spaces/philschmid/image-generation-editing">a HuggingFace wrapper</a> since Gemini 2.0 Flash is API-based. we're using a base prompt, <code class="inline-code">a photo of three goths eating pancakes at an IHOP</code>. Here's the first result we get:

Not a high-quality image, but it's a place to start! Now let's modify the image.

<code class="inline-code">make one of the goths dressed up like santa claus</code>

Sort of? One of them is dressed up like santa claus, but it's a totally different person. Everything in the image has been regenerated from scratch, as if behind the scenes it was just connecting the two prompts together and generating a new image. Let's see what happens when we change something else.

<code class="inline-code">make a different goth be dressed up for mardi gras</code>

This changed the santa part but in this case the rest of the image is largely un-changed. So clearly this is something the AI could do (edit just a part of the image). Let's keep going.&nbsp;

<code class="inline-code">make a different goth dressed up like a ghost for halloween</code>

Now we've got one goth sort of wearing a white shirt, and the santa is a robber for some reason? Nobody's sure what's going on with the left goth who's been mean-mugging the whole time.

<code class="inline-code">make the stack of pancakes so tall it goes out of frame</code>

Other than the fact that they look poorly-photoshopped-in, this did what the prompt asked. So at least there's that!

<code class="inline-code">make it so instead of eating pancakes they are eating frisbees</code>

We get a different, non-frisbee related stack, still out of frame. Can't consider this a success. Also each generation makes everything look more fake, like plastic.&nbsp;

Did we learn anything with this experiment? Some combination of multimodal interaction and Google's Gemini 2.0 Flash model are just not ready for the type of edits that marketers, or really anyone who wants to use these tools for anything meaningful, would want to do. We'll keep an eye on these tools as they evolve, but they're definitely not ready yet.

<code class="inline-code">the goths have finished eating all the pancakes are are waving goodbye</code>

Discover why Google's Gemini 2.0 Flash fails hilariously at editing AI-generated images of goths eating pancakes at IHOP. The "not quite" problem persists.