In-Between Frame Generation for 2D Animation Using Generative Adversarial Networks
Abstract
Traditional 2D animation remains a largely manual process where each frame in a video is hand-drawn, as no robust algorithmic solutions exist to assist in this process. This project introduces a system that generates intermediate frames in an uncolored 2D animated video sequence using Generative Adversarial Networks (GAN), a deep learning approach widely used for tasks within the creative realm. We treat the task as a frame interpolation problem, and show that adding a GAN dynamic to a system significantly improves the perceptual fidelity of the generated images, as measured by perceptual oriented metrics that aim to capture human judgment of image quality. Moreover, this thesis proposes a simple end-to-end training framework that avoids domain transferability issues that arise when leveraging components pre-trained on natural video. Lastly, we show that the two main challenges for frame interpolation in this domain, large motion and information sparsity, interact such that the magnitude of objects' motion across frames conditions the appearance of artifacts associated with information sparsity.
Subject Area
Computer science|Artificial intelligence
Recommended Citation
Arriaga Pazos, Francisco, "In-Between Frame Generation for 2D Animation Using Generative Adversarial Networks" (2024). ETD Collection for University of Texas, El Paso. AAI30820528.
https://scholarworks.utep.edu/dissertations/AAI30820528