Stelios Perrakis

I am currently pursuing an MSc in Computer Science 🎓 at the University of Crete. My thesis focuses on applying Generative Models in Deep Learning 💻 to address challenges in image super-resolution 🔍, employing advanced techniques such as Diffusion and Latent Diffusion Models. My primary interest lies in leveraging Deep Learning to tackle challenges in Computer Vision 👀.

Keywords: Generative Models, Deep Learning, Image super-resolution, Diffusion Models, Latent Diffusion Models, Computer Vision

news

Dec 10, 2024	My MSc thesis involved the exploration of diffusion models for super-resolution in medical imaging. Specifically Improved, Cold, and ResShift variants, with various backbones including UNet and VIT. These models were tested in both the Latent and Spatial domains, utilizing first-stage models trained in-house, such as VQVAE and VAE. Their performance was compared against established models from hugging face MRI Autoencoder v0.1 and the SDXL-VAE. In the in-house trained VAE, I incorporated a specialized training loss that enabled the model to converge in fewer iterations compared to the traditional combination of GAN loss, KL divergence, and pixel loss. Additionally, I experimented with LoRA to handle data from out-of-distribution data, such as different acceleration factors in MRI images and data variations from FastMRI knee to FastMRI brain and vice versa. I plan to open-source the code and will post extensive analyses on various aspects of this work. Stay tuned for more updates on each topic.
Dec 01, 2024	I am currently seeking a PhD position or a research role or an Internship, that aligns with my interests in computer vision. If you know of any opportunities or have any available, please do not hesitate to contact me. I am willing to start immediately.

Dec 10, 2024

My MSc thesis involved the exploration of diffusion models for super-resolution in medical imaging. Specifically Improved, Cold, and ResShift variants, with various backbones including UNet and VIT. These models were tested in both the Latent and Spatial domains, utilizing first-stage models trained in-house, such as VQVAE and VAE. Their performance was compared against established models from hugging face MRI Autoencoder v0.1 and the SDXL-VAE. In the in-house trained VAE, I incorporated a specialized training loss that enabled the model to converge in fewer iterations compared to the traditional combination of GAN loss, KL divergence, and pixel loss. Additionally, I experimented with LoRA to handle data from out-of-distribution data, such as different acceleration factors in MRI images and data variations from FastMRI knee to FastMRI brain and vice versa. I plan to open-source the code and will post extensive analyses on various aspects of this work. Stay tuned for more updates on each topic.

Dec 01, 2024

I am currently seeking a PhD position or a research role or an Internship, that aligns with my interests in computer vision. If you know of any opportunities or have any available, please do not hesitate to contact me. I am willing to start immediately.