Stable Diffusion is a versatile generative AI model that can be used to create images. Artists and designers can leverage these models to create a fun logo with little effort – we’ll show you how.
Table of Contents
Before We Begin
When testing several different versions of Stable Diffusion, we found that the base XL model provided the best results. The 1.x and 2.x models had a tough time creating good images that also required upscaling. Given the computing demands of the XL model, you may need to rely on a cloud provider or rental GPU service such as RunDiffusion, Vast.ai, etc. We used Vast.ai for our testing and the total cost was approximately $2 for all logos used in this article.
Related: Check out our comparison of Stable Diffusion models if you want to know why the XL model is so much more robust.
Overview of The Process
Here are a few things to keep in mind before generating a logo:
- Basic composition must be done by you before letting Stable Diffusion do its work.
- Most work will be done through the img2img endpoint.
- The process is very iterative.
- With each iteration, you need to lead the model to the result you are looking for. This means that you will need to modify an image, then take that modified image and modify it further.
- Don’t forget to save a .txt output for each generation.
Need additional guidance? Watch our video that walks you through the step-by-step process.
Creating a Logo for Single Letter or 2-3 Letters
The most primitive and easiest logo will be a stylized letter or short collection of letters.
To start, you’ll want to use any image or word processor and simply type the letter and general color that you’d like your final logo to represent. It’s okay if it isn’t exact – however, a general idea of the color and shape will drastically reduce the amount of work and effort needed from the model during the iterative process.
When setting up the letter and color, be sure that the final dimension of the image is 1024 x 1024 as this is the scale on which the XL model was trained.
For our example, we fired up Pixelmator and typed the letter “Ai” in a large font. We then changed the color to a dark blue.
The result is below:
Again, the software being used here isn’t important – only the canvas size, letter, and color matter. Plenty of online services can also be used.
In our example, we’ll be using Automatic 1111 WebUI to interact with the Stable Diffusion XL model. Here’s our guide if you need help setting it up.
Once the application is running, do the following steps:
- Click the img2img tab
- Upload your image
- Enter your prompt and negative prompt. Explicitly tell the model what you are currently uploading and some modifiers. For example, “blue letter C with a white background, melting, complementary highlights.” For the negative prompt at minimum add in “watermark, trademark, signature.” You can also add in other words that you don’t want to appear in the image as you continue to iterate.
- Denoising strength defaults to .75, however, you may want to adjust this as it will tell the model how closely it should adhere to the uploaded image (0 will result in the same image whereas 1 may be completely different). I have found anywhere between .51 and .85 to be the allowable range. CFG may also be changed if you want to force the model to follow your prompt (learn what the CFG scale actually does).
- While optional, setting the batch size to 4 allows you to evaluate several images at once so you can quickly decide which is on the right track for the result you are looking for.
- Once you have the above set, click the “Generate Image” button.
Depending on the GPU, this may take anywhere from a few seconds to a few minutes.
If the image is not what you are looking for, click the “Generate Image” button again. However, if one image is heading in the right direction, then click the painting portrait (below the output image):
Once you send the modified image to the img2img tab, you will then be iterating based on the modified image (not the original one you made). While you can hit the “generate image” button to generate a new image, you can also modify the prompt and negative prompt to help steer the model.
You’ll continue doing this until you get the final result you are looking for. In our example, we went about 5 levels deep and several iterations at each level before getting to the final result:
Stylizing a Word
Stable diffusion doesn’t perform well with text outputs. Therefore, with word logos, it’s best to temper expectations and expect slightly stylized outputs. If you want to do a play on the letters (for example, an “i” that looks like an ice cream cone), you will have to take a more creative approach.
Just like generating single-letter logos, it’s best to place your word into a 1024 x 1024 canvas with a white background.
When setting the prompt, be explicit “a photo of the word ice cream.” Then add in the modifiers you want to see.
As for the denoising strength, you’ll want to keep this lower to preserve the legibility of the final output. The CFG scale could also be raised slightly to ensure that the model more closely follows the prompt.
Transforming Reference Images
The process behind this is similar to stylizing a single letter. You’ll want to describe the image, then add appropriate modifiers to help direct the model to where you want it to go.
In our example, we had this clip art image of a fish where we wanted to direct it to a picture of a fishing lure. Therefore, we
Pre-pending a Logo to a Word
A lot of company branding involves a logo that is set to the left of the company name. For example, the Apple logo is set to the left of the word “Apple.” Doing this isn’t nearly as hard as the final results may suggest.
This approach requires two separate processes. First, let’s generate the logo:
- Generate a rough composition of a potential logo using txt2img. To make it easier use prompts like “an illustration of a coffee mug, white background”
- Set the composition or layout of the image in your preferred editor:
- Send the potential candidate to the img2img tab and refine further:
This will help to generate a more cohesive look between the logo and the word. Continue to iterate until you are satisfied with the output.
LoRA Models and ControlNet
The scope of this article is to generate a traditional and unique logo rather than one that is set on a background or in a specific stylized setting. With that being said, there are some LoRA models for logos on Civitai. However, the results tend to be rather spotty – the most notable is the Harrlogos LoRA.
ControlNet is another option, to consider exploring. Using the canny edge detection algorithm, ControlNet can be used to generate a logo that follows a typeface more closely.