Guide to Selecting Images for AI Model Training

Choosing the right images is key to training great AI models. Follow our guide to make the process easier and more effective!

Tip:

Be sure to also check out our full guide on using Model Trainer. It walks you through the entire process of creating effective Characters and Styles, beyond just photo selection!

Resolution & Aspect Ratio

Each image should have a resolution of at least 1 megapixel, roughly 1024x1024 pixels or higher.

❌ low-res image

✅ high-res image

We automatically crop images to the closest supported resolution, so any aspect ratio is okay. However, more aspect ratios don't automatically mean a better model.
It’s best to use a few aspect ratios, ideally the ones you’ll generate images in later. But don't worry—you can still generate great images in different ratios even if all your training images only have the same one (e.g., are all squares).

Quantity

To start training, you need at least 3 images. For optimal results, aim for 10 to 20 images.
More images mean longer training times, but often lead to better model performance.
Don’t upload the same image multiple times!

Image Content

Character Models

Avoid blurry or out-of-focus images, as they make it harder for the model to learn.
Make sure the Character is alone—no other people.
Use images where the Character is of a similar age. Mixing old and recent photos can confuse the model.
Don’t include photos of your Character wearing accessories like sunglasses, hats, or heavy makeup.

❌ wearing sunglasses

❌ too heavy make-up

❌ blurry photo

❌ not a solo picture

Use photos showing the Character with different expressions (smiling, laughing, serious, etc.), lighting, and backgrounds.
Match poses and shot types to your goal. For general character models, use a wide range of poses and shot types (close-ups, medium shots, and full-body shots). If your focus is on generating portraits, prioritize well-cropped portrait shots.

Style Models

All images should reflect the same aesthetic. Include images of different objects, people, and scenes to give the model a good sense of the Style’s applications.

✅ Diverse training set

Consistency is REALLY important. Even a single image that deviates from the chosen Style can disrupt the model’s ability to generate correct images with similar subjects. For example, a single picture of a dog drawn in a different aesthetic can cause all animals not to be generated in the target Style.

❌ Incorrect training set

(image in different style)

Tip:

Your model finished training and is marked as "Active"? Check out our guide on using your Characters & Styles and see it in action!

Follow these tips, and you'll be well on your way to creating amazing AI models. Happy training!