Okay, so allow me to start with a quick demo of Zyro’s AI-generated logo visuals changing to music:
Trust me. It will all make more sense soon. But first, a little about how Zyro’s AI Logo Generator was born.
Zyro AI Logo Generator: The origin story
As with a lot of what we do, our idea for an AI Logo Generator began with a team discussion. We always want to stay on top of all the newest and coolest AI and machine learning trends out there, including Generative Adversarial Networks (GANs) and what they can do.
Which, as it turns out, is actually quite a lot. If you don’t believe us, here are some top examples of GANs at work.
Pro tip: refresh the pages to see new realistic creations.
If you’re still hungry for more, you can find plenty more awesome GAN creations, but by now, you can probably tell that GANs can generate some pretty realistic-looking stuff.
Honestly, one example that really stood out from the rest came from an unlikely source at Waifu Labs.
Waifu Labs was remarkable for a few reasons. Not only were users able to generate unique characters, but they could also influence the outcome with just a few simple, visual steps.
In practice, this meant that people could have a fun, interactive, and effective experience generating AI images.
Better yet, it was easy, and users could skip all the boring steps. And okay, animated characters aren’t much use to us at Zyro, but that’s when it clicked: what if we could do the same for logos?
The rest, as they say, is history. Zyro’s AI-powered Logo Generator had been born, and now we just had to make it work.
Implementation and training hurdles
Initially, we thought that it wouldn’t be too hard to create an AI Logo Generator. Just find an existing model, or, at the very least, a dataset to train a model, pop it into the production, and that’s it, right?
Well, as it happens, the answer to that question is no.
The first paper and implementation we found on the topic did not have the results we wanted. We did not have the resources to build our own model, and it didn’t look like anyone else had found what we were looking for either (spoiler alert: some researchers were already working on it).
And honestly, that was a real blocker for us for quite some time. Well, at least until the StyleGAN2 paper was published and the results inspired us to try it out and train it ourselves.
First things first, we needed data. The existing datasets weren’t good enough, and so we decided to gather our own. We used a combination of publically available logos from various sources on the internet, and finally, it was time to start training the StyleGAN2 model.
Training began on Google Cloud Platform, but we quickly noticed that the model needed to train for weeks, which would cost us a small fortune.
Still, we knew that there were other solutions that we could use, and one of those was FluidStack.io.
After the initial training, we decided to check out the results. Obviously, we were thrilled that we had achieved something, but the overall progress still wasn’t what we strive to create at Zyro.
See for yourself:
Okay, let’s zoom in a little for a better look.
Let’s be honest. It’s absolutely full of unreadable gibberish. Only 1 in around 15 images even come anywhere close to looking like a legitimate logo.
We have to fix it somehow, but how?
After a little bit of research, we decided to try out a semi-supervised learning technique with a simple neural network. The idea was to weed out the textual logos from our dataset so that the model wouldn’t try to replicate them going forward.
So, after filtering out the data and retraining the model, we finally got rid of the gibberish text, and these were our results:
Now, let’s zoom in once more.
But wait. What is this? What are these creatures with such uncanny faces?
Taking a thorough look through the dataset, we noticed that it includes some images of dolls and kids’ toys. Of course, the model learned this and then tried to replicate it.
We must go back and clean these dolls out of our dataset. Obviously, we didn’t want to train the model from scratch, as it is pretty costly and takes a lot of time, so we decided to further train the same model based on our new dataset.
Here is the result:
Much better. Sure, we still have pandas and a few other humanoid figures, but these were taken from actual logos and now only represent a minority of the output.
Still, for Zyro, this isn’t enough. The output looks a little messy, and many of the logos do not look strictly professional.
We thought long and hard about how to solve the problem until we concluded that adding some vectorization could make the logos flatter and more modern while also allowing users to scale them without any pixelation.
And we were right. It worked! See the results for yourself:
The vectorization method solved most of the issues, and now we can finally generate some good-looking logos.
The new model also allows us to quickly refresh the generated logo selection so that we can choose some beautiful logos from a massive set of options.
AI Logo Generator future vision
Of course, this output is still not perfect, but this is Zyro, and we never stop trying to improve.
Right now, we already have some ideas brewing, including:
- Free control over which letter the final logo resembles
- Allowing the user to choose an initial logo style from a set of predefines
- Continued improvements to the output that will further enhance logo realism
As always, feel free to share your ideas, offer feedback and suggestions, and it may just appear in the next version of Zyro’s AI Logo Generator.