Can Generative AI create content for rubber.tube?

03/10/2023

Let's face it, it would make things much easier if I could outsource content to AI. But it it good enough to make me stuff yet? And what better time to do this that now, when everyone is no longer excited or impressed by AI image generation and have pretty much gotten over the whole thing? Lets start by trying to get it to reproduce the classic Tim Credible cartoon.

The classic toon is detailed with lots of items all over the horse. Tim Credible is cool and smug. The horse "Trampy" doesn't seem too concerned that his legs have been folded in half. Overall it's a pretty stupid drawing. Can AI draw something like this? Let's have some attempts from various AI models.

The classic Tim Credible Toon, circa 2006

First attempts

I started using Bing Image Creator a few months back, when it was still using a modified Dall·E 2 model. Results were funny, but not particularly amazing.

I loved the vibe of this one:

Prompt: "A too-cool-for-school knight on a fat horse with short legs". The one legged horse result was very funny. I don't know what the words on the horse was trying to say, "lol cat" or something.

And proceeded to animate it...

The animated DALL·E 2 image. The one legged horse was really quie a funny.

Craiyon

Craiyon is a much more lightweight model compared to the others, and it shows. It sucks pretty bad at this sort of thing. The best I could manage was this image, but nothing really interesting here.

Craiyon attempt. Prompt: a cool knight riding a comically fat horse, cartoon style

Photoshop Generative AI

Photoshop recently added generative AI to it's beta version. The advantage is you can generate a whole image, or select regions of the image and have it generate content to add to the image. As such, adding detailed items can be done one at a time, such as generating the knight on a horse first, then adding the details.

The first photoshop Generative AI attempt. The main scene was generated first, then I added in some of the details such as the castle, the maiden, the gramophone etc.

I could get most things to work, alkthough some things like the cannon were troublesome as it has restictions on 'offensive' or 'violent' imagery which cannon was flagged as. Also the lowered horse legs didn't seem to work well. Here's a particularly entertaining attempt.

An attempt at 'lowering' the horse legs. The prompt was "horse legs folded in half with padded bottoms" or variations on this theme.
The second photoshop Generative AI attempt. In a similar way I created the main scene first, then tweaked aspects and added details.

DALL·E 3 (Bing image creator)

Recently Bing updated it to use the newer dalle3 model, so I thought I'd give this one a try. Starting with a descriptive prompt for a cartoon image, we end up with some lovely images. I liked this one the most, although it's missing some of the description, it's a decent and cohesive image. I noticed that most generations seemed to lack drawing a cannon or cannonette. Sometimes, like this one, the cannon is in the background, but never was it attached to the saddle as requested.

prompt: a cartoon of a cool knight on a fat horse. the horse has short legs. the knight has shiny armour and a red cape. in the background is a castle or medieval structures. on the back of a horse is a gramophone. hooked onto the saddle of the horse is a small cannon. the horse has a fluffy dice around it's neck.

With a cartoon working pretty well, I thought I'd try some photorealitic generations. For some reason, it had a hard time making it look like a real knight, and kept going for a "toy" photo look. They were pretty cool, but not what I was after.

prompt: a photorealstic image of a cool medieval knight outside in the countryside. The knight is riding a fat horse. The knight and horse are real and lifesize. the horse has short legs. the knight has shiny armour and a red cape. The knight's visor only covers his eyes and rests on his nose in the background is a castle or medieval structures. on the back of a horse is a gramophone. hooked onto the saddle of the horse is a small cannon. the horse has a fluffy dice around it's neck.

Altering the prompt iteratively, I came up with some nice ones, although perhaps not as photorealitic as I would have liked. This one kind of looks like a 3d render, or maybe a model.

prompt: a photorealstic image of a cool medieval knight outside in the countryside. The knight is riding a fat horse. The knight and horse are real and lifesize. the horse has short legs. the knight has shiny armour and a red cape. The knight's visor only covers his eyes and rests on his nose in the background is a castle or medieval structures. on the back of a horse is a gramophone. hooked onto the saddle of the horse is a small cannon. Sigma 24 mm f/8

Using the same prompt but with more generations, another one I liked. I do find it interesting how often it puts the cape on the horse.

prompt: a photorealstic image of a cool medieval knight outside in the countryside. The knight is riding a fat horse. The knight and horse are real and lifesize. the horse has short legs. the knight has shiny armour and a red cape. The knight's visor only covers his eyes and rests on his nose in the background is a castle or medieval structures. on the back of a horse is a gramophone. hooked onto the saddle of the horse is a small cannon. Sigma 24 mm f/8

I changed styles here, and told Bing about a cartoon character I've created and all of the details. With the new DALL·E 3 updates, Bing will keep track of subsequent updates to the character or image as you suggest more changes.

A cartoon style first try.

Interesting that it decided to start labelling things. It did well with the red feather and the gramophone. I suggested some changes.

A cartoon style v2.

Some of these are quite hilarious, I love the style and the fact the it's putting sunglasses over the helmet.

A storybook cartoon style.

For the final attempt I suggested a more children's storybook style. I was really happy with this result, and I left it there for Bing's DALL·E 3.

The final image, I'm really happy with this one!

Midjourney

I got some mixed results from Midjourney. It certainly excels in portrait photo styles, and has trouble creating exactly what I told it.

First midjourney result for "cool knight on a fat horse with short legs", some very dramatic scenes, like game concept art.

I thought I'd lean into the realistic photo for now. These images are amazing. You'd be hard pressed to tell they aren't real photos! Adding camera lenses to the prompt helps it create a photorealisic image.

Changing the prompt to include "photo of" and "Sigma 24 mm f/8" made it photorealistic.

Not bad, he needs a helmet, but I had trouble getting it to only partially cover his face, so opted for medieval sunglasses.

Adding "wearing medieval sunglasses" makes a cool looking knight. Not exactly Tim Credible, but what would I know, perhaps this is exactly what Tim Credible would be like.

It was looking good enough, I upscaled one of the images I thought was my favourite.

Upscaled image

Midjourney 5 has a new "zoom out" feature so you can generate more of the image.

Zoomed out image to see more of Trampy.

I tried to do a more cartoon style, although it resulted in a dramatic style.

prompt: "Cartoon of a cool knight on a comically fat horse"

To make the style more cartoonish, I needed to give it a "style of", which Gary Larson seemed to produce good results.

prompt: "Cartoon of a cool knight on a comically fat horse in the style of Gary Larson"

Probably the best one I could make. Not bad, but missing all the details. Overall this is pretty impressive for an AI model.

prompt: "Cartoon of a cool knight on a comically fat horse in the style of Gary Larson"

Conclusions

It was fun experimenting with different generative AI models to see what I could get, and to understand it's limitations. You can get some pretty impressive results, it's just hard to get it to do exactly what you want. So can it create content I can publish on the site? Well it just did. But boy was it a lot of work...