tl;dr - Prompt shouldn't be too specific and shouldn't have too many subjects. Add your own styles and/or upload composition image to have more control.

Two generation models are available: Creative and Structured. Structured model is used by default. Structured model is much better at animals, faces, characters and overall structure of the generated image. It also has less repetition. The model lacks details and does not work well with long prompts.

Creative model is much better at details and can handle long prompts relatively well. It works amazingly well with landscape and abstract prompts. This model is not great for animals, faces, characters and many times subject gets cut randomly.

This guide has comparison examples from both models.

Since generation is random and we can only control it via words, Snowpixel generates a few images for a single prompt to increase likelihood of having at least one good generation.

In most cases you need to hone the focus down to two to three subjects and a goal of how those interact. Model has been trained on internet data till December 2020. If subject had lots of images on internet at the time, better chance it will be generated well.

One word prompts allow for abstraction by the AI whereas a small sentence has a more focused lens. You have to ask yourself if you are looking for something specific or abstract:


A one-word prompt so simple can allow the AI to go anywhere it likes, the emotion, the people or the objects associated with it.

The comfort of a bedroom, cozy and clean, a lovely bed with furniture

A more specific simple prompt, much less abstract outputs and much cleaner looks.

Make sure you are not being too specific, give AI some leeway to be creative. Kanye West winning 5 Nobel Prizes at Old Trafford Stadium is very specific. Whereas A person winning prize at a football stadium is not and helps AI to be creative. For the same reason, abstract phrases like Intimacy of having conversations with oneself work well.

Poetic Prompting: Using expressive language often times results in pretty good generations. Some examples:

It is beautiful here, I can feel the presence of the spirit of this place

2 subjects, the Viewer, and the place its effect on the emotion of the viewer and how the place interacts with the viewer are all in this simple line. These are the results:

The past is forever, the future is always waiting, this photo is timeless

3 subjects with abstract ideas attached to them, this combination leads to more abstract outputs.

As you can see in above examples, it can get quite hard to control AI on where to place subjects. You can have more control over this by using a composition Image. This guide goes into more details.

Snowpixel generates images in a few different styles by default. Results generated from prompt: Willow tree with Creative model.

As you can see, all sorts of styles and mediums are here like monochrome, sketch, illustrations and paintings.

You can also generate multiple images in a single style. Here are some results generated by locking in Jackson Pollock style, sketch medium and purple color. Modified Prompt: Sketch of purple willow tree by Jackson Pollock with Creative model.

You can design a prompt by adding a medium prefix, a style suffix or both.

Style Suffix

With a suffix, the modified prompt will look like

  • Willow tree by Van Gogh
  • Willow tree in fauvism style
  • Willow tree in glowing neon colors on purple background
  • Willow tree by Da Vinci made of lego blocks
  • Willow tree by any random name

It does not have to be a known artist. It could be any name, try using your own and see what it generates! If you generate dozens of images with the same random name, you'll notice that there's an associated style with it.

You can get very good results by choosing an artist compatible with the prompt. Work of artist Ilya Kuvshinov includes lots of faces, if you use Willow tree by Ilya Kuvshinov, there's a good chance a face will pop up somewhere. Similarly, if your prompt is Intergalactic Warfare, choosing an artist whose works have elements of Space Art, Science fiction like Vincent Di Fate will increase likelihood of getting these elements in your generations.

Wikiart has a good list of styles, genres and mediums.

Medium Prefix

With a prefix, the modified prompt would look like

  • Watercolor painting of Willow tree
  • Glitch art of Willow tree
  • Sketch of Willow tree
  • Pixel art of Willow tree
  • Mixed media artwork of Willow tree

Medium prefix doesn't work as well on Structured model. It's recommended to skip it there.

Combining medium and styles

  • Ink wash painting of Willow tree by Andy Warhol in vibrant colors
  • Dry brush watercolor painting of Willow tree in pastel colors
  • Concept art of Willow tree by Eyvind Earle in minimalism style

If you are not getting good generations consistently, please let us know at and we'll help you out with designing better prompts.