Pokémon Card Generator

Input image 1
Generated Pokémon 1 (Dark type)
Final Card 1

My group will NOT be presenting. However, DevPost makes the checkbox a required option to submit the project, so we had to check the box.**

Inspiration

With HackTX's emphasis on AI, as well as the ever-growing hype around it, my team wanted to work on a project that involved some sort of AI.

So, we brainstormed several ideas that could utilize AI and combine some of our personal interests! We wanted to make something cool, but also have fun while doing it! As a result, we came up with this project!

What it does

This project asks the user to input an image and a card typing. Then, we produce a Pokémon card based on that input.

How we built it

Used Python to run the main code

Challenges we ran into

Initially, we wanted to build a model from the ground up. However, we found that there was a lack of datasets that paired Pokémon to their card. Furthermore, there are a bunch of cards for the same Pokémon. And training on this data might not provide the results we wanted in the time we had.
- As a result, we went for a less ambitious idea that offloaded more of the work from AI and gave it to humans (us).
Our computers were slow and we had limited time to work, so we weren't able to run the best models
- Additionally, we have to think of the end-user. Their devices might take too long to run super big models. As a result, we chose to go with smaller, faster models, sacrificing some of the model's performance.

Accomplishments that we're proud of

First and foremost, we're proud of all of our abilities to collaborate and synergize each of our strengths to create a product we're proud of.
We're also proud of

What we learned

We learned some of the limitations of AI. It's not something that we can manipulate very easily to do whatever we want. Instead, we must carefully engineer the model and data to achieve specific tasks. There is a lot of human process into designing these systems with AI.
We also learned how important data and training time are.

What's next for Pokémon Card Generator

With more time to craft the data and train models, we would like to either train a model from the ground up to work specifically with our Pokémon to Pokémon card dataset OR fine-tune an existing image-to-image model on our dataset (rather than just utilizing the base, pre-trained model). We believe that if we do this, we can get the model to generate the entire card, rather than just the Pokémon image.

We would also like to make the interface more friendly to the end user.

Stay tuned!!!