Training an in-game item style in Layer helps you generate assets like trophies, gems, weapons, tools, or collectibles - all with a consistent visual look. Whether your items are designed for RPGs, puzzle games, or match-3 gameplay, this guide will help you train a style that delivers clean, on-model results.
Start with Your Assets
Begin by preparing your training dataset. Use high-quality images of in-game items that follow the same structure and perspective. If you’re working with isometric assets, make sure every item is aligned to the same angle.
We recommend using around 25 images for best results. Try to keep the styling consistent across every asset - that includes lighting direction, stroke thickness, color palette, shadow treatment, and rendering style.
Why Captions Matter (and What to Watch For)
When you upload your assets, Layer automatically creates captions for each image to guide the training. These are a good starting point - but for item styles, it’s worth reviewing every caption closely.
Things to look out for:
Is the perspective or view accurate? For example, if all your assets are “isometric view of a [item],” make sure that phrasing is consistent across captions.
Are key style traits called out? If all your items have thick outlines, glowing edges, or a flat color style, include those details in every caption.
Do the captions feel uniform? Consistent structure across all your captions helps the model better understand the pattern of your style.
If you’re planning to use the same format for every caption — such as “An isometric view of a golden trophy, stylized 2D art with soft shadows and thick outlines” — consider saving that for use as a Prompt Prefix + Suffix after training.
How to Start Training
Once your captions are reviewed and updated, you’re ready to train your style.
Follow the steps in the How to Create a Custom Style guide:
 | Go to the Art Styles page and click Build a New Style |
 | Upload your assets or create a style from selected files in your Layer Drive |
 | Choose a Style Base (SDXL, BRIA, or FLUX) |
 | Select In-Game Items as your Style Type |
 | Add at least 3 example prompts to show what kinds of items you want to generate |
 | Start training — you’ll be notified when it’s complete |
Best Practices for In-Game Item Style Training
 | Keep perspective consistent |
 | Style traits matter |
 | Avoid mixing styles |
 | Match your dataset to your goals |
While It’s Training: Set Up Prompt Prefix + Suffix
Once training is underway, go to your style’s settings and set up a Prompt Prefix and Suffix to reinforce structure during generation.
For example:
Prompt Prefix: “An isometric view of a [item name]”
Prompt Suffix: “Stylized 2D art with thick black outlines, soft shadows, and a glowing rim light”
This ensures your prompts always start and end with consistent styling language, keeping your outputs on-model.