How We Make Images

Get a glimpse into how we make AI images!

We are a free stock image website, but every image is created via AI and then the best images are hand-picked. The process we use to make the images is described below:

What is Stable Diffusion

Before we begin, we should describe Stable Diffusion. In extremely simplified terms, Stable Diffusion can be thought of as a reverse captioning system. You provide an image caption (typically called a prompt), and it creates an image that roughly matches it. The model is improving all the time, but as of writing this, it can sometimes be tricky to get it to generate exactly what you want. In our experience, we get 'good' images ~5-10% of the time.

We personally use the crowd-sourced machines made available via the AI Horde.

Step 1: Writing a Prompt

We have a system in place to randomly create prompts given various lists of subjects, styles, adjectives, etc. It combines these to create random prompts such as a cute panda, oil painting or something more complex like a caveman holding a torch in a cave, art in the style of Van Gogh.

We're also always adding new terms to use in the random prompts to add variety.

Step 2: Generate the Image

The next step is to generate the image! We input the prompt to the Stable Diffusion model and it spits out an image file!

Step 3: Manual Review

We are always slowly generating new pictures. However, many of the generated pictures end up being low quality. There are many situations that can cause a picture to be 'bad':

• A bad or nonsensical prompt. This usually happens when our system picks keywords that just don't make sense together. We try our best to limit this possibility, but some slip through occasionally.
• Bad parameters to the model. We also randomize many of the parameters we use when generating pictures. The details aren't important here, but sometimes certain combinations of parameters can result in less-than-ideal images.
• Bad luck. In our experience, even with a great prompt, the output image won't always come out looking high quality. It might be blurry, or missing detail, or have ugly image artifacts. We commonly see animals with multiple tails, people with blurry faces, etc.
• Personal preference. We also may decide an image is "good" by most measures, but still decide it isn't useful to publish.
• Real People or Brands. Sometimes a picture will include a real person or a brand logo. We always avoid these. For example, images of sneakers commonly include a Nike logo and images of apples sometimes render as the Apple logo.

Step 4: Editing

Some images are good but have minor flaws or blemishes that would be easy to fix with image editing software. We do that if the editing needed is minor. This is usually just removing spots that didn't look good, or cropping an image a bit.

Step 5: Publishing!

Once an image has been reviewed and deemed high quality, we publish it! Once published, the images immediately go live on the site for our users!