Step 4: Training Dataset

The Dataset tab is the central library that your model learns from. Every image you upload, character poses, facial expressions, action shots, and environmental scenes, collectively forms the knowledge base the AI uses to understand your IP. The stronger and more balanced your dataset is, the more accurate and consistent your outputs will be.

A key part of building a reliable dataset is including both character images and environment images. Characters alone aren’t enough. The model also needs to see the world your character lives in: locations, lighting, textures, color palettes, and the general mood of your universe. When the dataset includes a healthy mix of both, the AI learns how your character interacts with different spaces, which significantly reduces the chance of hallucination or distorted outputs later.

---

As you upload images, the system automatically generates captions describing five core attributes of each photo:

• Pose

• Emotion

• Body proportions (full body, upper body, etc.)

• Camera angle

• Action

These captions help the model correctly identify what it’s looking at. If any of the captions are inaccurate, you can edit them directly. Once your edits are complete, regenerate captions so the entire dataset stays synchronized and accurate.Making small corrections here ensures that the model isn’t trained on misunderstandings or mislabeled images.

One of the powerful parts of the Dataset tab is that it supports continuous learning. Your dataset isn’t static, it can grow with your project. Any time you generate an image you like, you can add it back into your dataset. Over time, this continuous cycle strengthens the model’s understanding of your character’s style, range, and environment. The model essentially evolves with you, becoming more accurate the more you use it.

---

How It Works In Summary

After uploading your character images:

1. The model processes each photo individually.

2. It generates an automatic caption (description) for each one.

3. These captions may include: -“standing, full body”, “three-quarter view” , “side profile”, “surprised”, “front facing”, “reaching” etc.

4. The reviewing user reviews and edits any incorrect captions.

5. Updated captions are saved and pushed back into the dataset.