Skip to content

Latest commit

 

History

History
25 lines (16 loc) · 2.29 KB

README.md

File metadata and controls

25 lines (16 loc) · 2.29 KB

Frankenstein Network (FranNet)

There is a live demo of the project available here, which allows you to upload your own images and see the results of the network. However, I highly recommend reading the documentation first to understand the limitations of the project. Also, hit "Restart" in case the demo is was stopped due to inactivity.

FranNet is a neural network architecture that combines a CNN encoder with a NeRF-like decoder to perform upscaling, denoising, and colorization of input images. The purpose of this project is to explore the capabilities and limitations of this approach and implement it as a proof of concept. A bit more detailed description of the network can be found in the overview.

The choice of upscaling and colorization tasks was made because they are interesting and visually appealing, and they do not require a significant amount of resources. Additionally, these tasks lend themselves well to the application of NeRF (Neural Radiance Fields). While NeRF is typically used for different purposes, I was intrigued by its ability to generate images from individual rays/points, which is ideal for upscaling.

DISCLAIMER/WARNING: The purpose of this project is primarily exploratory, and it is not feasible to achieve any form of photorealism due to severe resource limitations. As a result, the practical applicability of the project is currently limited. The neural network utilized in this project has been designed with fewer than 600,000 parameters, leading to a reduction in overall quality. It is important to note that all experiments were constrained to 15 training epochs (2-3 hours on Google Colab using Tesla T4 GPU).

Project documentation navigation

Technical details of the project:

Additionally, there are some Useful Links to articles, videos, and other resources that were beneficial during the project.