ULTRA OFFICES – JUST BEFORE WAITLIST LAUNCH

You’re launching a new men’s grooming brand, but you need a new logo. The one you’ve been using so far is just some placeholder text in a nice font. However you don’t have the time or money to hire a designer, what do you do?

Ok we need a logo and it has to look decent enough
Otherwise people won’t sign up to our waitlist
We could probably spare a few thousand dollars from the budget…
But really it’s the time we can’t afford: the last branding project I worked on took months
So for now why don’t you try and generate one using one of these AI tools?
Stable Diffusion is really blowing up on my Twitter feed
The quality looks as good as DALL-E but it’s free
Check it out and see what you can come up with

Stable Diffusion is the latest in a quick succession of releases of AI art tools. DALL-E by OpenAI kicked off interest in the space and showed everyone what was possible. Then DALL-E 2 came out in beta and it was a step change in quality: good enough to actually replace a human designer on some tasks for the first time. Unexpectedly given that DALL-E 2 looked like magic, competitors quickly arose in Midjourney and Craiyon (initially called DALL-E Mini), which grew to millions of members from all the people who couldn’t get DALL-E access. Early adopters posted amazing photos generated by AI on social media, which drove even more interest in this space, once people could see for themselves what was suddenly possible.
Then right after OpenAI released pricing for DALL-E 2, Stable Diffusion was released, a completely open-source and free model, “made for the people, by the people”, as it said on the stability.ai website. The sudden emergence of a free tool you can run in your browser with just a little coding knowledge has exploded interest in the industry, leading to a flourishing of innovation and creativity that promises to completely upend the creative industries. People are learning ‘prompt engineering’ to figure out how to create good images, and building tools that allow even more advanced uses of these AI models.
Most people still aren’t aware of just how good these tools can be, or how accessible they are to use, so the shift won’t happen overnight. But as an early adopter of these tools you can learn how to work with these tools to solve design problems like creating a new logo, saving you thousands of dollars on hiring a human designer. Even if you do go down the human path, having an almost finished product to show them that matches your vision is infinitely better than trying to describe what you want, and takes 10x less work for both sides.

Stable Diffusion

Hey, you're going to walk you through how to create a company logo with stable deficient for free in your Google collab notebook.
So this is the step of diffusion Google CoLab. They set this up so you can kind of run this in the Google environment.
You don't need to, you know, download anything onto your computer. So it's much easier to run. And it also gives you a free GPU, which is like a specialized computer.
That's like very good at running this type of thing quickly. So we're just going to click, create a copy to drive and then that's going to pretty copy.
It's just creating a copy right now. Might take a couple of minutes. Okay. A copy is created and we're basically interested in go through this setup process him.
So now we have this copy. We can call it whatever I want. It's this size now. And we can run this it says a required tire Ram.
Don't worry about that. Okay. Just come in and get rid of any of the, you know, the sessions, whatever. So hopefully this will run now to connect.
Okay.  So that is connecting just to make sure that you're doing this for a way you can just go to change, run time type.
So you see this hardware accelerator GPU that's that's what you need to make sure is set to GP. Okay. Now we're connected and says, we've got this computer set up here.
Okay. School. So now we can run this. Okay. That's run fine. And then we're just going to go through it and run these intent.
So as this is running, you might want to set up a hugging face an account. So in order to do this you need to go open this model card up here and you can see I'm logged in a talking face.
But if you're, if you're not logged in you'll see something like this,  We'll see the middle card. And then once you sign up, it will kind of ask you to approve that mobile card.
So once you've done that then then you should be good in order to run it. Okay. So, so that's run.
Okay. And then we're going to create this she's going to run this as that lets you kind of customize the output and then this is going to lock you into hugging face.
Okay. So once you have a hugging face account, you can click on that and then see I have this token and you can click copy token or if you need to create a new one, just name it and change it to write and generate the token.
But I've got that one here. So I can copy that and then paste it in there and then click login.
And that will give you this error, but you don't need to worry about that. If, if you get this era then you know, it's, it's working cool.
So now we're just need to download all the kind of model stuff. Don't worry too much about what's happening here.
It's basically downloading this table diffusion model from hugging face which is kind of like a you know, platform that like holds all these models and data sets that you can use.
So lots of really cool stuff on there and you know, it's free to use as well. So so really, really good way to get started with AI.
Okay. So this is downloading the next thing we're going to have to do once it's downloaded is set up this pipe.
So this is basically how it runs everything on the GPU rather than like on a normal computer.  And like I said before, GPS are kind of optimized to do AI really well.
And normally they're very expensive to buy like thousands of dollars. So really cool, like Google offices for free, although you know, they are starting to try and get people to upgrade to the pro version.
So I don't know how much longer this will be free. Okay. So it's downloading all this stuff here. I'm just going to take a little bit of time trying to pause it while it's telling me okay.
That only took a few minutes to download and then run this Pope. So when you're running this, since you need click play or what I'm doing is pressing shift and, and to when I'm selected in this code so now we ready to create our first image.
And in order to create an image, we just need a few things. One is the prompt. So this is a prompt that gave you a photograph of an astronaut riding, a horse and then it uses this with auto casting to basically run this on the GPU hour or on faster.
And then it's going to grab the image and then save the image or you can you know, display it here because we're in Google collapse.
So let's try this and see what we get. So what is doing is doing diffusion. So the way this works is it's starting with a completely random noisy image.
And then it's kind of running a, an encoded that that basically kind of de noises the picture. So like if you have a restored, an old photograph it de noises it, and then it's trying to basically the noise, it in a way that it matches this prompt.
So it has like another piece that says, does this match a photograph of a national running horse? And if it doesn't then will keep kind of moving in that direction until you get something like this.
Cool. So actually, you know, works, we have a photograph of a national riding horse and you can do this in Dolly which is the open AI version.
But because several fusion is, is open source, it's free pretty much to run it. So you know, obviously you can pay more for more compute power.
If you throw more compute power at it, the images get better. But it's pretty cool to just try it out in here.
And then if you do want to use a paid version, that's maybe a bit more accurate than you could use a Dolly after you've kind of figured out what props work.
So now in this section, when you use a manual seed and what manual see does, is it meat? This sort of starts the prompter like starts to noisy images, certain kind of level of random noise.
And that means that if anyone who runs the same seed and the same prompt, it will get the exact same image every time.
So before you saw that image of Nashville riding a horse or as like this one came up with a new image and we ran it, this one should bring us back the exact same image.
So this is really useful. Yeah, here we go. So exact same image. So this is really useful for reproducibility. And all you need to do is kind of set the generator here and then push the generator in to the pipe like that.
Cool. The other thing we can do is kind of choose how many steps. So like, I think the default is a hundred, but you could set it to be less or more.
So this is still setting that manual seed. Like we said before, it's the exact same image but we're doing it with less steps.
And that's where you can kind of see you know, just how compute makes a big difference to the problem.  It also takes less time to run if you have less steps.
So I play around with that and kind of get an intuition for you know, how much kind of competing power is needed to get the quality that you need.
Cool. So the other thing we're going to do here is just generate three different images at once as this is just a function that they've given you to create a grid.
And then when you run it you can get three different images now. So it's basically running it on to three different images and just gonna come out the other side and you know, put three different images out there.
So this is what you kind of expect from Dali or magenty or any of the other crayons, if you've used that because they, they give you like a great of images and this is doing the same thing.
Just kinda pause while that's running. Cool. So that took 40 seconds to run and you can see some are much better than others.
Like, you know, this looks like a video game, you know, this doesn't have a horse in it, but this one looks great.
Actually, that's a really cool picture. Cool. And then you can create like, as many images as you want. Like this is a running it, I guess, for 12 images.
So quite a few. So you know, that's going to take a long time to run, probably not going to do that.
So you know, and, and there's obviously the more you can do this tutorial keeps going on and tells you how to change the height and width if you want.
These have to be divisible by eight just to like, you know, you're running into an error otherwise.  It also explains a lot more about how the model works, which is interesting.
But we had to do a Trump, which is to create a new company Loker. So let's do that. And we're gonna take this and we're just gonna we're gonna do a copy of the cell and then piece this out again, that'd be important torturing more when we're not committed to do this generator sort of not gonna set around the seed.
Actually let me turn around and see if she can get the same area. Just me, if, if you want to use it for some reason cool.
And number of inference steps, we're going to deluxe up to 200 because we want to get like a pretty good quality.
Okay. So I found it, I probably emotionally can get it to like two 50. I think when I do 300, it starts to complainant me.
I need to use a bigger computer, probably. Cool. So what do we want? So we have a shaving brand called ultra and it's a men's grooming brand.
They do cheering foam deodorant, and I want to create a new logo for them. So you know, we can say a new library for a new company for ultra and we could see a, well that comes back with C yeah.
Okay. That only took a minute to run and we got a pretty cool logo actually. One thing you'll notice is that the words never make any sense unless you're using tons and tons of info steps.
This is just a, a problem with the model in general, as you get the same and Dolly as well. But like, that's not what we want.
We want this logo here and what you can do is you know, now you can kind of save the image if you want.
So we can take that curd up here, you know, image safe. And then let's, if you liked, this could say and then in your files is gonna show up very Oh, here we go.
No, it's that cool. And then you can just download that and you get it on your local computer. And then what I think this is really good for is just ideation.
Like, you'll probably, won't be the final version, although there, you could use that. It depends on how you know, how early stage your brand is, like how, how much quality you need.
But you can kind of drop this into Photoshop or Figma, kind of edit out the words, and then you got like the little bit that you need and kind of work on it from there.
So the really cool thing here is that you, as a, non-technical like non-design non creative person you can actually create your logo for free and, and, you know, try lots of different styles until you find what you like.
And then and then go use the Cuban, right? Like it's going to be much easier brief for the designer. If you just say, this is what I like, can you make like a better version of this?
Then that's going to save them doing tons and tons of legwork, just trying to second, guess what you want, because now you can visually show them the style that you need.
So should bring down the cost of working with the designer by, you know, 10 X. All right. Hopefully that was useful.


Why might it be a good idea to create your logo with AI?

What do you need to do first to use Stable Diffusion?

What’s the downside of using Stable Diffusion in Google Colab Notebook?

That’s it – I’m not hiring designers to make logos for any of my new projects. It used to cost thousands of dollars every time I spun up a new venture, but now I make logos for free in 15 minutes using AI. The results are actually *better* than what I got before.

Vexpower Learning Certificate of Completion

Congratulations fabulous-drone-53 💪