How to make a logo with Stable Diffusion

It’s easier than you think. Let’s make this golf cart logo.

Lesson Goals

  • Browse AI art models suitable for logos
  • Create a logo using different art styles
  • Learn about the limitations of text and AI


Not bad, but we can do better. It’s a little busy.

Quick and Dirty method

The logo above was created with this prompt:

/render #sdxl an illustrated logo of a golf cart on a golf course, award-winning logo, masterpiece iconic artwork [[bad quality, low quality, jpg, signatures, watermarks]]

This uses the prompt recipe #sdxl and some popular keywords to boost quality.  This is a totally valid way to do it, but you can get more control by digging into the models system.


Step 1: Find Logo Models

You can search right from Telegram to look for base art styles, like this:

/concept /search:logo

Some ideas will appear as a list.  We also provide a web page companion with graphics, of course, but this list is fast and handy too.

At a glance, it looks like there’s quite a few LoRAs (mini models) focused on logos, like <logo-modern> and <logo-redmond-xl>

Let’s see what they look like in the web browser by typing /concepts and clicking the mini browser.

See the graphical version instead: type /concepts to browse by tag or pop up the visual models list

Step 2: Let’s pick a Logos LoRA

These are the generations of Stable Diffusions, entire groups of AI models that use a different technology. As such, the models within each family are kept separate in our software, as they are not compatible with each other.

Here we are switching to the SDXL models to see the models we saw earlier in the text list.


Step 3:  Copy the Trigger word

In the Telegram List and in the browser, you can easily find the trigger word.

Switching to different art styles

If you look at the image above, the MuteMan and Ci
nsdia offer slightly different logo illustration styles. So if we don’t like one style, you can try another.  Here’s the same prompt using Cinsdia’s more 2D flat stylings

You can also change the Base Model, the underlying art style, from SDXL to another XL-based base model. This means you can mix and match LoRAs and Base Models to give that LoRA a whole new dimension.

I’m looking for an illustration, so a base model like DreamShaper XL seems like a popular choice. We’re constantly upgrading the models, so pay attention to the trigger name as a new version might be available. At this time, it is <dreamshaper1-xl>

LoRAs also have Weights, the amount of influence they have on the image. You can increase or decrease the influence of a LoRA, which helps with its compatibility as well. In fact, you can theoretically stack multiple LoRAs to get your desired effect.

Put it all together

/render #sdxl an illustrated logo of a golf cart on a golf course, award-winning logo, masterpiece iconic artwork [[bad quality, low quality, jpg, signatures, watermarks]]<logo-redmond> <dreamshaper1-xl>

Same exact prompt, different supporting art styles and models, and that’s so much closer to what I wanted.  A cute logo in seconds. You can now click on Tools to upscale it or remove the background in one click, those things are also included with our tools. There is no additional cost for upscaling, our tool is unlimited.


How about Text?

There are some models that are better at text than others. In fact, if you add the Lora <bettertext-xl> from the XL family, it may handle your text with a little more clarity.

BetterText-XL works best by adding /size:1024×1024 to your prompt, which is the exact native XL resolution.

But please do curb your expectations on text. Prompting can only get you so far with generating text with AI.

These are the early days of Artificial Intelligence, and text has not been solved with 100% accuracy. With the exception of a few short well known words, it will struggle with even basic lettering. Prompting for letters in a logo isn’t the best way to do it, unless the words are absolutely short and simple.

There is a better way.

ControlNet is the best AI Text solution

There is a 100% accurate, absolute best way to do text — and that is to create the text template outline first in another application, then upload it into our ControlNet AI tool.


Fine Tuning Logo Techniques

Prompt Guidance

Guidance, also known as CFG, is how literal the AI should follow your prompt.

Sure, you said “jump”, but how high?

Guidance is a number between 0 and 20.  Low guidance allows the AI to be more creative, but the image may get fuzzy. Choosing a high guidance will make the AI more confident about following your prompt down to the letter, but may introduce glitches if there are conflicts with the LoRAs or keywords.

A guidance around 7-10 is the most common.

You can also use shorthand command as part of your positive prompt, like: /guidance:20


Size and Aspect Ratio

The default aspect ratio is Square, but it has the highest compatibility. Easy presets are provided in the advanced menu.  You can also manually specify a size, but remember that AI models are trained on specific preset sizes, so specifying a gigantic resolution requires more video ram, and may result in duplicate objects.

Don’t try to get to 4K in one shot, even on 48GB of Video RAM it isn’t enough to process that file unless you want your logo by tomorrow. Instead, you can use the UPSCALE and FACELIFT to get to or beyond 4K.

A safe range for SD 1.5 models is as follows:


A safe range for SD XL models starts at 1024×1024 and goes to about 1300×1300 before you run into problems.

Positive and negatives

Make your prompt more specific by adding positive weights to words that matter like ((black and white)) and use negative weights to things you want to remove like [[trees, background, orange, fuzzy textures, green, blue]]

Switch up your prompt

You can also naming specific colors, brands, or artists that you like. Any text change will result in a slightly different image. Re-running the exact same prompt will also give you more variations, but click on the Tools > More button will save you time.


back to tutorials