Cartoon style generation of animated image scenes based on improved GAN and RetNet
- 1Graduate School, Cheongju University, Heongju, Korea
- 2School of Design and Art, Shandong Huayu University of Technology, Dezhou, China
Abstract
To improve the quality and style consistency of cartoon generation in animated image scenes, this study proposes an image cartoon generation model that integrates improved generative adversarial networks and preservation networks. The model achieves collaborative optimization of local details, global structure, and diverse styles in images by introducing a multi-style mapping module, a multi-head discriminator, a Euclidean distance decay attention mechanism, and a thin plate spline geometric consistency module. This study employs two publicly available scene image datasets, setting up four typical simulation testing environments and comparing them with three mainstream image generation methods. The experiment showed that the model performed well in key indicators such as generation quality, processing efficiency, and structural preservation, with a peak signal-to-noise ratio of up to 29.2 dB, the Fréchet Inception Distance was 35.2, and a processing efficiency of 8.0 FPS. In addition, the generated cartoon image had a structural similarity index of 0.89 with the original image, and the running memory usage was controlled within 3.9GB. The image generation model has overcome the style drift and geometric distortion in existing animated cartoonization, providing a high-fidelity and low-cost automated solution for film and television production, game development, and digital content creation.
Related
- Brain-Aligned Multimodal LLMs for Social and Emotional Understanding of Cartoon Movies
- Tailoring capsid-directed evolution technology for improved AAV-mediated CAR-T generation
- Fixed-ground multisensor monitoring resolves longitudinal vegetation condition and fuel-state structure in Southern California chaparral and oak scenes
- Shifts in neural representations, not gaze, track how people reinterpret ambiguous social scenes
- The Identification of Biological Stains at Crime Scenes: A Promising Role for Proteomics and Machine Learning
BibTeX
@article{jin2026cartoon,
title = {{Cartoon style generation of animated image scenes based on improved GAN and RetNet}},
author = {Zhenyao Jin and Lei Yang},
journal = {PLOS One},
year = {2026},
month = sep,
doi = {10.1371/journal.pone.0347014},
url = {https://doi.org/10.1371/journal.pone.0347014},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Cartoon style generation of animated image scenes based on improved GAN and RetNet
AU - Zhenyao Jin
AU - Lei Yang
PY - 2026
DA - 2026/09/24
JO - PLOS One
DO - 10.1371/journal.pone.0347014
UR - https://doi.org/10.1371/journal.pone.0347014
AB - To improve the quality and style consistency of cartoon generation in animated image scenes, this study proposes an image cartoon generation model that integrates improved generative adversarial networks and preservation networks. The model achieves collaborative optimization of local details, global structure, and diverse styles in images by introducing a multi-style mapping module, a multi-head discriminator, a Euclidean distance decay attention mechanism, and a thin plate spline geometric consistency module. This study employs two publicly available scene image datasets, setting up four typical simulation testing environments and comparing them with three mainstream image generation methods. The experiment showed that the model performed well in key indicators such as generation quality, processing efficiency, and structural preservation, with a peak signal-to-noise ratio of up to 29.2 dB, the Fréchet Inception Distance was 35.2, and a processing efficiency of 8.0 FPS. In addition, the generated cartoon image had a structural similarity index of 0.89 with the original image, and the running memory usage was controlled within 3.9GB. The image generation model has overcome the style drift and geometric distortion in existing animated cartoonization, providing a high-fidelity and low-cost automated solution for film and television production, game development, and digital content creation.
SN - 1932-6203
ER -