How to create your own sticker pack using the Stable Diffusion neural network

Did you know that you don't have to be a designer to create your own beautiful sticker pack?

Today I will tell the story of the creation of mine sticker pack from 20+ stickers With mine character named Evelynn using neural networks. You don't need a powerful video card or even Photoshop(In most cases): I'll show you how to do all this directly in the browser. This is a simple, accessible and free way to create unique stickers, even a kettle can handle it, I hope it's already boiled, let's get started!

Making stickers is a creative endeavor, and methods, which I will show, Can easily combineto speed up and simplify processes.
I'll start by telling you about all the necessary tools, and then I'll move on to the process of creating stickers, showing each step with examples. You'll see how easy it is to bring your ideas to life using the Stable Diffusion neural network and available online services.

Why not Midjourney? Because in Stable Diffusion you can more flexible control the final result of generation, on the go correct the shortcomingssuch as fingers or any details. But recently it has been possible right there get rid of backgroundwithout leaving the interface. If you are afraid of SD, then it is in vain, because it is not complicated at all.

ArtGeneration

If I didn't have a video card and needed a cool online image generation service, then this would definitely be artgeneration.me.
Just recently they implemented the ability to add LoRA (these are small models for specific tasks) and brought in my favorite XL model – animapencil XL.

The service is free, but there is a daily limit on generations. But I made it for you referral linkaccording to which you will receive week of PRO accountalthough even after a week the functionality is sufficient for us.

RunDiffusion

Generative AI in the cloud. The service is paid, from 50 rubles per hour. Personally, I have not had time to use it yet, but there is a choice: Automatic1111 or Fooocus. If you are not familiar with the Automatic1111 interface or have never encountered image generation before, then I recommend you choose Fooocus if you decide to go with this service. There is no difference between them, except for the interface, and Fooocus has a much clearer one.

Fooocus locally

If you are a lucky owner of an NVIDIA video card with more than 8 GB of video memory, then congratulations, you can install the Fooocus interface locally!

To install the original Fooocus, go to link on GitHub and click on >>> Click here to download <<<.

Unpack the archive into any convenient folder without Cyrillic in the name.

Next we see the following picture:

Don't go into details, just run run.bat. The necessary files will be downloaded and Fooocus will be ready to work.

Forge Interface

If you are an experienced user, you can use Forge, especially since now it is possible to generate images on a transparent background. I have not tried it with stickers, but if you are interested, I recommend you take a look Forge video from my friend Nerual Dreming. I will tell and show everything on the Fooocus interface, and if you are well versed in Forge, you can easily adapt my advice for it.

Everything I will tell you further can be applied to each of the services.

Preparing for work

At this stage we make blanks for future stickers. I use the model animagineXLbut later tried on animapencilXL and I like both of them very much. Well, from LoRA we will only need Redmond Stickers.

Oh yeah, I forgot to tell you what's what!

If you already know the Fooocus interface, then proceed to step 1.

Models or checkpoints – trained neural network models. The models have an understanding of the appearance of a wide range of objects, such as people, fauna, buildings, and more.

I divide models into realistic, cartoon, anime models and everything else.

LoRA – these are small models. If large checkpoints are needed to draw a wide range of objects, then LoRA is needed for specific objects, styles, characters, etc.

That's very briefly. Just repeat after me and over time you'll understand what is responsible for what.

Above I have listed the models that we will use. CivitAI you will be able to find all sorts of models for every taste and color. I know that you will want to experiment and use something different, so pay attention to the Base Model item. You need models with SDXL 1.0 architecture.

We download models and LoRA and send them to the root folder Fooocus>models>checkpoints or loras.

That's it! Let's restart the interface via run.bat.

The following window will open in your browser:

We press Advanced and a menu appears on the right. This is what we need.

On the left bottom we see the Prompt window and the gray Generate button. In the window we enter the request, i.e. what we want to see on the image.

A quick tour of the interface:

All we need from the first tab at first is:

Resolution – as the name suggests, the resolution of your image. I recommend setting it to 1024×1024 and not worrying about it.
Image Number – the number of output images. Optimally 4.
Negative Prompt – what should not be in the image is entered into this window.

Styles. For stickers, uncheck all the boxes, they will only get in the way.

The models and LoRA that we already know. We select what we downloaded and set it up as in the screenshot (you can move it around and set it up to your taste, but these are the best parameters that I have deduced).

In the Advanced tab, set the following parameters. The rest is not needed.

We've briefly gone over the interface, but if you give it a lot of pluses, I'll tell you more about the interface, and which parameter is responsible for what. Within the framework of this guide, it will all take a very long time.

Step 1. Generate stickers

Stickers should have a simple prompt (text request). But you should put the most important features of the character and his appearance in it. Here's what I used:

sticker, simple background, pretty beautiful 1girl with short white hair, hair buns, white t-shirt, freckles, blue eyes, __Emotion, pose__

At the end, I always included a description of the emotion or pose, which is the most important thing in the sticker. Below I have collected what I used, I had to sort through, some poses or emotions in English are called differently than I expected:

Emotions:

light smile / light smirk – smiling / with a smirk

Smirk – an evil grin

angry, irritated, – angry, irritated

surprised expression, amazed, open mouth – surprised

one eye closed, winking, – winking

serious – serious

crying, sad – sad, with tears

embarassed, confused, – embarrassed

focused on – focused on an object

Gestures:

thumb up – “like” gesture

showing thumb with his hand – points a finger at the viewer

waving her/his palm – waving her hand, greeting

heart shaped hands (you will need AwyHandHeartXL, weight 0.8) – heart shaped hands

saluting – saluting

hands on head – holds head

shrugging – to shrug one's shoulders

I often had to google sites with the names of anime emotions:

The point is that you shouldn't always trust the translator, I often additionally google the translation of a word or expression into English. Well, I use DeepL as a translator.

Let's start.

sticker, simple background, pretty beautiful 1girl with short white hair, hair buns, white t-shirt, freckles, blue eyes, waving her palm

The negative query I used was:

bad quality, bad anatomy, worst quality, poor quality, monochrome, low quality, low resolution, blurry, blur, watermark, lowres ugly, disfigured, duplicate, mutated, bad art, blur, blurry, dof

The resolution I mostly used was either 1024×1024 or 768×1280.

LoRA weight – 1.4.

The most successful is the fourth picture. It does not need to be expanded, the background turned out white, as I intended. We will use the fourth, sometimes the background may not turn out as in the first options, in this case just increase the weight of the white background (select the white background in the prompt and press ctrl + up arrow).

Next, I improve the eyes and face, removing artifacts using the Inpainting tool.

I drag the image into the window and draw a mask, first selecting the eyes.

Inpainting has three modes:

Inpaint – standard mode. Most often I use it to correct hands/remove artifacts, etc.

Improve – a mode for increasing the detail of individual parts of the image (face, eyes, hands, clothes, etc.)

Modify – used to add objects, change background, etc.

This is in general terms, in fact, each uses a different denoising strength and inpaint model. But in this article we will not go into too much detail.

I select the eyes. I recommend using hotkeys in the Inpainting window:

Most often I use ready-made template hints below.

We generate.

Left – eyes after Improve Inpainting – more details and saturation.

Now let's remove all unnecessary things and artifacts, switch to the classic Inpaint mode.

Here's what we get in the end: I don't do Upscale, because Telegram itself has a limitation on uploading stickers of 512×512 pixels. Instead, I do Improve Inpaint on the eyes, face, and clothes one by one. Later in the article, I'll show how I added a print of my logo to a T-shirt.

Let's give a couple more examples and move on to the next step.

sticker, simple background, pretty beautiful 1girl with short white hair, saluting, one eye closed, hair buns, white t-shirt, freckles, blue eyes

The first option is the best one. Let's work on it.

First, let's use Outpaint – expand the image to the right. To do this, switch to Inpaint and select the classic Inpaint or Outtaint Method. And select the expansion to the left.

Great, now the left hand is completely visible. We do the same: clean the artifacts and then Improve Inpaint.

Lifehack: if a print from a T-shirt or something else just won't go away, try leaving the Prompt field empty, in this case it helped

Here I tried the following query using the “shy fingers” construct:

sticker, white background, pretty beautiful 1girl with short white hair, shy fingers, confused, hair buns, white t-shirt, freckles, blue eyes

Or let's make a sad emotion:

I think the approach is clear. The main thing is a bright emotion that we achieve through a request.

Let's move on to the next step.

Step 2. Finalizing the sticker

First, you need to remove the background. This can be done in two ways.

Method 1. For the lazy.

We just use online services, here are the ones I used:

I'll say right away that the edges aren't always smooth, and the hair on female characters isn't always cut out correctly. But often everything is clear, and after outlining it won't be particularly noticeable.
But, for nerds like me, there is a second way. And it flows smoothly into the next micro-stage.

Load our future sticker into Photoshop. Double-click to remove the lock in the layers.

Next, take the mouse and select the “Pen” tool.

And we start to outline our tian. It is “a little” longer than removing the background through an online service, but it is a classic and reliable method.

Also, in the latest versions of Photoshop there is a function “Remove background”, I noticed it only now, it works quite well, all that remains is to clean up the artifacts:

But I will show you what to do if you have an older version.

After you have selected the outline, click on the selection. Don't worry if you have small “white” areas left, they can be cleaned up in the same way.

I usually set the blending to 3-4, but you can also adjust it by eye using “Refine Edge”.

And there will be a panel on the right. We configure everything to taste.

Next, simply return to the image and press Ctrl + X to cut out the tangle.

In the “Layers” window, add a new one and move the character onto it. Done, you've gotten rid of the background!

We cut out the remaining pieces using the pen and selection in the same way.

The next step is adding a stroke. It can also be done online using various services, but I will continue working in Photoshop.

Right-click on the image layer – blending options

We need a stroke and a shadow.

We make a white outline on the outside, I advise you to make the same outline size on all stickers, just remember the number.

It is not necessary to make a very noticeable shadow, this one is quite enough

The text remains. It depends on the situation. If it is needed in the sticker, add it. I used this this one a selection of popular meme fonts. You can use any to your taste or not use it.

I also add a small shadow and a black outline to the text for readability.

Next, we need to adjust the resolution so that Telegram accepts the sticker.

The resolution is 512×512 and the maximum size is 512 KB.

In Photoshop, click Image – Image Size. Set to 512×512.

He promised to show how to place a print on a T-shirt.

1. Using a pen, outline the item (in this case, a T-shirt). Not the entire thing, but the area where the print will be.

2. Select the area.

3. Next, use “Refine Edge” and sliders to taste.

4. Copy the selected item to a new layer and remove the stroke and shadow effects (click on the eye):

5. CTRL + J. Duplicate the layer with the T-shirt.

6. CTRL + G. Create a group from the copied layer (click on it first):

7. Click on the layer in the group and create a “Levels” adjustment layer:

8. Click the bottom button, which I highlighted with an oval, and adjust the top sliders, increasing the contrast. Make it so that the maximum number of folds are visible.

9. Switch to the bottom layer with the T-shirt and transfer the logo. Right mouse button – create a clipping mask

10. Switch to the group and select the blending mode “Screen” or “multiply”, whichever is better.

We get the following result. Now Evelynn has the average salary in Russia. The logo looks more natural, highlights are visible, it is better than just transferring the image and leaving it as is.