Bouguila, M. Khaled (2024) Conditional Generative Modeling for Layout-Transforming PFE - Project Graduation, ENSTA.
![]()
| PDF 6Mb |
Abstract
This report presents the work conducted during a research internship at Criteo, focusing on Conditional Generative Modeling for Layout Transformation, a novel task in the field of image-to-image translation. Traditional models, such as CycleGAN, often struggle with significant layout differences between input and output images, as they tend to converge on mappings that retain similar overall structures. This work investigates methods to overcome this challenge by employing state-of-the-art generative models, namely Diffusion Models and MaskGIT. To tackle layout transformation, the study introduces novel approaches including the creation of a synthetic dataset and a new evaluation metric tailored to this task. Through extensive experimentation and architectural improvements, the results demonstrate both the strengths and limitations of current models, proposing new directions for latent space design to better handle diverse layouts. This work lays the foundation for future research in improving generative models for complex layout transformations.
Item Type: | Thesis (PFE - Project Graduation) |
---|---|
Uncontrolled Keywords: | Conditional Generative Modeling, Layout Transformation, Image-to-Image Translation, Diffusion Models, MaskGIT, Generative AI, Latent Space Design |
Subjects: | Mathematics and Applications |
ID Code: | 10432 |
Deposited By: | Khaled BOUGUILA |
Deposited On: | 11 oct. 2024 19:50 |
Dernière modification: | 11 oct. 2024 19:50 |
Repository Staff Only: item control page