Developing visually consistent game assets is crucial for maintaining the aesthetic integrity of a game. This guide provides a structured approach to training custom style AI models on Scenario, allowing game developers to generate assets that consistently reflect their unique visual style. Beginning with the selection of representative images and accurate captioning, the workflow progresses through the training of style models using Scenario. This method ensures that every element, from backgrounds to character designs, adheres to a unified artistic vision, enhancing the overall gameplay experience. By adhering to this training protocol, developers can achieve a high degree of stylistic consistency across all game levels and assets.
You can find numerous examples of Styles in the Platform Models section on Scenario - accessible through Models > Platform Models. We encourage you to explore and experiment with these pre-trained models to get a feel for the diverse range of styles available on the platform!
Even with the Starter (free) Plan, anyone can train their own styles on Scenario. Trainings are unlimited for paid subscribers (Creator and above), enabling you to experiment freely and discover the perfect models and styles for your project.
Begin by navigating to the + New Model button located in the Models section of the app. After selecting Start Training, follow this workflow:
For a deeper dive, continue to the steps below.
The images used in training are crucial for overall success. A typical training dataset for a style usually consists of 10 to 25 images, although it is possible to achieve good results with a dataset outside this range.
When selecting images for training, ensure they are high-quality with a minimum resolution of 1024px. Using smaller images may negatively impact your training results. Your images should be diverse enough to capture the essence of the style you aim to achieve. Additionally, all images must have a 1:1 ratio square format, which you can crop directly in the web app.
Training images can be hand-drawn, digitally created, AI-generated, or a combination of these. You can also train directly in the web app using a collection or a series of generations. Simply select multiple images and click … > Train a New Model.
More advice on curating your dataset.
Captioning, or labeling, plays a vital role in helping the AI comprehend the content and style of the training images. Well-written and accurate captions contribute to more consistent and visually appealing outputs.
Effective captions are concise yet descriptive, vividly depicting the image as if explaining it to someone who cannot see it. Focus on the subject matter, composition, colors, and other relevant elements. However, avoid explicitly mentioning the artistic style itself in the captions, such as "anime," "watercolor," or "3D rendered." Allow the visual characteristics to speak for themselves, enabling the AI to interpret the style based on the provided examples.
Although Scenario offers an automated captioning option, it's highly recommended to review and adjust the captions so the reflect what you see in the image. When reviewing the captions, check for:
Before initiating the training process, select the Style Preset. It’s recommended to stick with the default training presets. However, if you're an advanced user looking to refine a style model, you can adjust these settings for more specific results.
Hit Train to begin the process. The training time will vary depending on the number of training images you've provided and can take anywhere from 20 minutes to a few hours.
Information on advanced training techniques.
Once the training is complete, begin by testing your model with simple prompts that resemble the original caption structure. If your captions consisted of only a few words per image, avoid starting with lengthy prompts. This approach will help you evaluate how well your model responds and the quality of its outputs.
Experiment with various prompts to find the optimal structure for your model. Keep a record of these prompts and consider including them in your model's description on Scenario for future reference. If necessary, you can also add negative prompts as a final step to further refine the outputs.
Developing the skill to quickly refine a model may require some practice to achieve the desired results. If your model needs improvement, you can easily retrain it by doing the following:
Another effective technique for refining your style is to compose your model by blending it with others. This approach allows you to combine the strengths of different models, creating a unique and more refined style. For instance:
Composing models unlocks a wealth of possibilities, enabling you to create custom styles tailored to your needs. To learn more about composing models and achieving the best results, explore our in-depth guide.
Style training is a crucial skill for users to develop when creating images with AI. If your initial training doesn't yield the desired results, we recommend experimenting with different combinations of images and captions to fine-tune your model. With practice, you'll become more proficient in achieving your desired style.