Animator2D

Animator2D is an artificial intelligence model designed to generate pixel-art sprite animations based on textual descriptions. The model uses a BERT-based text encoder to extract textual features and a convolutional generative network to create animated sprites.

Model Description

• Name: Animator2D
• Input:
  • Character description
  • Number of animation frames
  • Character action
  • Viewing direction
• Output: Animated sprite in image format

Dataset

The model was trained using the spraix_1024 dataset, which contains animated sprites with detailed textual descriptions.

Future Goals

This is only the first version of the model. In the future, we aim to improve it with the following updates:

• Expand output formats: Currently, the model generates a single frame sheet. We plan to implement the ability to export output in multiple formats, including folders containing separate images, animated GIFs, and videos.
• Optimize frame management: The current frame count is manually defined, but we aim to improve control by introducing a more intuitive system that considers factors such as FPS and the actual animation duration.
• Enhance the model: The current model is still in an early stage. Future updates will focus on making sprite generation more precise and consistent by improving architecture and training data quality.
• Customization of sprite dimensions: We will implement an input that allows specifying the character's height in pixels. This will enable adaptation of the generated sprite's graphical style, ensuring greater flexibility and customization possibilities (e.g., Pokémon style vs. Metal Slug style).