Getting started: Top tips to creating a successful AI model

On Exactly.ai, you can train an AI model to produce new work in your style, on your terms. Simply upload and caption a selection of 10 images or so, and around one hour later you will be able to prompt a personalised AI model to create new images. Plus, you can licence these models to others while retaining legal ownership of the work.

Sounds pretty easy, right? The process is simple, but we do have three top tips (plus a few for the pros!) that will ensure the best results every time.

Let’s get started.

Tip one: Quality and quantity make a difference

Your artwork is the most essential ingredient to a successful AI model, so let’s make sure it meets the requirements.

Firstly, size. In principle, bigger is better. Why? Because it gives the AI model plenty of information to learn from. JPEG or PNG files are best, and if you work on paper, scan or take high-res photos for the best quality. If you only have small images, make sure the shortest size is at least 512 pixels. For the upper limit, around 4000 pixels max, or 50MB for a single upload.

For quantity, we recommend at least 5 images, but ideally around 10. More images tend to produce better models, but that’s not always the case. We’ve seen plenty of successful models with just five or six reference images, so feel free to experiment.

Lastly, remember that your AI model will learn from the images you upload. If you have frames around your art, or fingers on the edges of your photographs, your model will include frames and fingers. For now, our program doesn’t generate text, so make sure to erase any watermarks or text inside speech bubbles so your model doesn’t get confused.

Tip two: Feed it diverse subjects, in a consistent style

Think of your AI model as The Hungry Caterpillar. You need to feed your caterpillar a variety of high-quality, nutrient-rich food in order for it to transform into a beautiful butterfly.

A successful model will create images in a strikingly similar style to the artwork you upload, but generally there are two key elements to achieving this: uploading a range of subjects, in a consistent visual style.

Let's tackle diversity first. It’s important to provide reference images that show different subjects. For example, if you upload ten drawings of a sunflower, your AI will only be good at drawing sunflowers. If you supplemented these with a person, dog, furniture and landscape, your AI model will be capable of producing a whole lot more with a closer likeness to your style.

Next, consistency. A consistent set of images uses the same technique with a similar colour palette. Your images should look like they belong to an artist series, or adhere to a specific brand’s guidelines. Any images that feel incongruent may confuse your model.

Tip three: Keep your captions simple and succinct

Once you’ve uploaded your images, Exactly.ai will ask you to caption them. This step is crucial, because these captions will later inform your AI model on how to interpret the prompts. Here’s a basic list of dos and don’ts.

Do:

• Write brief and consistent captions.
• Describe the image, not the image set.
• Use nouns, like a girl, two cats, or a table.
• Use the present tense; add ‘ing’ to any verbs, like ‘a girl running’

Unless you’re deliberately experimenting, don’t:

• Define the medium – painting, sculpture, or cartoon, for example – you’ve already provided the medium through your image set.
• Use adjectives that describe the style or mood of the image. For example, don’t write ‘a mystical illustration of a fairy’. Just write ‘a fairy’.

Your model will generate the best results when you prompt it using a similar format to the captions. Keep this in mind when you’re using someone else’s model too: if you want to create a watercolour landscape or an oil-painted portrait, for example, pick a model with those themes.

For now, that’s everything you need to know to get started. Remember, rules are made to be broken. Some of the most interesting models have intentionally avoided these rules to unlock new possibilities. Experiment with all the different ingredients yourself, until you find the recipe that suits you best.

Sign up now to train your own model.

Up Next: How to edit generated images.