Medievilizer

• 

  UI design

Inspiration

Throughout our first year, there was a master's student who was always there to encourage us throughout the year. He was always quite keen on the usage of computer vision, and the benefits it can provide - as he himself was producing amazing research in the field. As a result, we were inspired to bring the theme to life - by converting modern images into intriguing, unique medieval ones.

What it does

This project functions using a pre-trained stable diffusion model, which has been fine-tuned for generating medieval imagery using the LoRA method to convert modern images into medieval-styled ones.

How we built it

A generic diffusion model was customized for medieval-style generation via LoRA-based fine-tuning.

Challenges we ran into

There were a multitude of issues we encountered. For example, the time constraints inhibited us from training the diffusion model ourselves; which may have possibly improved our results. Furthermore, the lack of computing power made it challenging to produce satisfactory results in a short period of time - a large amount of clever engineering was needed to circumvent this.

Accomplishments that we're proud of

We initially had no idea on what even entails the process of image generation, and the progress from completely being clueless as to how to tackle such a problem, to the ending result was extremely fulfilling.

What we learned

Front-end development can indeed be quite challenging at times; but equally we now completely understand how diffusion models and neural networks work together to produce the stunning AI art we often see.

What's next for Mediev

We may further develop the front-end; or implement the original video-to-video conversion pipeline, which we had to scrap due to a lack of computing power.

Built With

• css
• diffusers
• flask
• html
• javascript
• lora
• python
• pytorch