Yusuf Dalva

I am a Ph.D. student at the department of computer science at Virginia Tech, Blacksburg, VA, United States. I am working on computer vision and machine learning, mainly focusing on generative models, with an emphasis on diffusion models. Before joining Virginia Tech, I obtained my M.Sc. & B.Sc. degree from Deprtament of Computer Engineering at Bilkent University (M.Sc. Thesis).

At Virginia Tech, I am currently working on the controllability aspect of generative models under the supervison of Pinar Yanardag. In the past, I have been fortunate to work with Aysegul Dundar at Bilkent University, on my work targeting image editing with GANs.

Until today, I worked on various topic related to computer vision such as:

Robustness enhancements of deep neural networks
Controlaability in generative models
Image-to-image translation on global semantics
Texture estimation for 3D meshes
Inversion methods for pre-trained GANs

Email / Resume / Bio / Google Scholar / Twitter / Github

Research

I'm interested in the intersection of computer vision and deep learning. My research primarily involves controlling the generative process offered by different image generation frameworks. A list of publications is provided below. Please see my Google Scholar page for an up-to-date list.

	NoiseCLR: A Contrastive Learning Approach for Unsupervised Discovery of Interpretable Directions in Diffusion Models Yusuf Dalva, Pinar Yanardag CVPR 2024 / bibtex (Oral - top 0.8%) We propose a contrastive learning based approach to discover latent directions on diffusion domains. We deomonstrate the effectiveness of our framework on multiple domains such as face images, cats, cars and artistic paintings. Paper / Project Page
	Image-to-Image Translation with Disentangled Latent Vectors for Face Editing Yusuf Dalva, Hamza Pehlivan, Oyku Irmak Hatipoglu, Cansu Moran, Aysegul Dundar IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) 2023 / bibtex Empowered by a attention-based feature selection and global embeddings representing images, we perform image-to-image translation using 6 different semantics. With the attention mechanism, we introduce filtering on facial features. Paper / Project Page
	Benchmarking the Robustness of Instance Segmentation Models Yusuf Dalva, Hamza Pehlivan, Said Fahri Altindis, Aysegul Dundar IEEE Transactions on Neural Networks and Learning Systems (TNNLS) 2023 / bibtex Presenting a comprehensive survey about the robustness of instance segmentation models. Evaluating state-of-the-art models comprehensively with both image corruptions, and cross-dataset benchmarking Paper
	StyleRes: Transforming the Residuals for Real Image Editing with StyleGAN Hamza Pehlivan, Yusuf Dalva, Aysegul Dundar CVPR 2023 / bibtex Proposing an image-to-image translation framework with an encoder-decoder architecture and latent vectors. Structuring the image space by learning these components altogether. Paper / Project Page
	VecGAN: Image-to-Image Translation with Interpretable Latent Directions Yusuf Dalva, Said Fahri Altindis, Aysegul Dundar ECCV 2022 / bibtex Proposing an image-to-image translation framework with an encoder-decoder architecture and latent vectors. Structuring the image space by learning these components altogether. Paper / Project Page

Teaching

CS 5914: Al Tools for Software Delivery (Spring 2024), Virginia Tech
CS 485/585: Deep Generative Networks (Spring 2023), Bilkent University
CS 464: Introduction to Machine Learning (Spring 2022, Fall 2022), Bilkent University
CS 342: Operating Systems (Spring 2022, Fall 2022, Spring 2023), Bilkent University
CS 101: Algorithms and Programming I (Fall 2021), Bilkent University
CS 224: Computer Organization (Fall 2020, Spring 2021), Bilkent University

Elected "Outstanding Teaching Assistant" in 2021, 2022, and 2023.

Adapted from the website of Jon Barron