Pengembangan Model AI NijiGAN menggunakan Contrasive Learning dan Persamaan Diferensiasi untuk Translasi Gambar Nyata ke Anime

Fadhila, Farah Dhia and Setyawan, Dimas Prihady (2025) Pengembangan Model AI NijiGAN menggunakan Contrasive Learning dan Persamaan Diferensiasi untuk Translasi Gambar Nyata ke Anime. Project Report. [s.n.]. (Unpublished)

Text Laporan KP_5025211030_5025211184.pdf - Accepted Version Download (895kB)

Abstract

Generative AI telah merevolusi industri animasi, termasuk konversi gambar dunia nyata menjadi anime melalui translasi tanpa pasangan. Scenimefy, model berbasis contrastive learning, mencapai hasil fidelitas tinggi namun terbatas oleh ketergantungan pada data berpasangan berkualitas rendah dan arsitektur berparameter tinggi. Penelitian ini memperkenalkan NijiGAN, model inovatif yang memanfaatkan Neural Ordinary Differential Equations (NeuralODEs) untuk transformasi berkelanjutan. NijiGAN menggunakan separuh parameter Scenimefy dengan data pseudo-paired untuk pelatihan, sehingga meningkatkan proses pelatihan dan menghilangkan kebutuhan data berkualitas rendah. Evaluasi menunjukkan NijiGAN menghasilkan gambar berkualitas lebih tinggi dengan Mean Opinion Score (MOS) 2,192, melampaui AnimeGAN (2,160), serta skor Frechet Inception Distance (FID) 58,71, lebih baik dari Scenimefy (60,32). Hasil ini menegaskan bahwa NijiGAN kompetitif sebagai alternatif state-of-the-art dalam translasi gambar.
==================================================================================================================================
Generative AI has revolutionized the animation industry, including the conversion of real-world images into anime through unsupervised image-to-image translation. Scenimefy, a model based on contrastive learning, achieves high-fidelity results but is limited by its reliance on low-quality paired data and a high-parameter architecture. This study introduces NijiGAN, an innovative model leveraging Neural Ordinary Differential Equations (NeuralODEs) for continuous transformations. NijiGAN operates with half the parameters of Scenimefy and utilizes pseudo-paired data for training, streamlining the training process and eliminating the dependence on low-quality data. Evaluations demonstrate that NijiGAN produces higher-quality images, achieving a Mean Opinion Score (MOS) of 2.192, surpassing AnimeGAN (2.160), and a Frechet Inception Distance (FID) score of 58.71, outperforming Scenimefy (60.32). These results confirm that NijiGAN is a competitive alternative to state-of-the-art models for image-to-image translation.

Item Type:	Monograph (Project Report)
Uncontrolled Keywords:	Image-to-Image Translation, NeuralODEs, Semi-supervised Training, Generative Artificial Intelligence, Translasi Gambar ke Gambar, NeuralODEs, Pelatihan Semi-supervised, Generative Artificial Intelligence.
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Dimas Prihady Setyawan
Date Deposited:	03 Jan 2025 12:39
Last Modified:	03 Jan 2025 12:39
URI:	http://repository.its.ac.id/id/eprint/116141

Actions (login required)

View Item