TelePhoto | Devpost

Prompt to create new image
Game mode selection
Gallery of past images

Inspiration

Do you remember the Telephone Game back in elementary school? Is there a way to connect this with modern generative AI techniques to bring a new dimension to this game? Combining these intriguing questions with the OUTER SPACE theme, we are inspired to bring Telephoto to reality.

What it does

The core idea is simple. Instead of manually whispering to the person next to you, we encode the secret text into an image and pass it along. After your guess on what the image represents, a gallery will be presented with the history of all past players before you. You can see how accurate you are, how the storyline changes, or even how far you went from the first guess! There are no limits to how you decide to play around with Telephoto.

How we built it

We used the Next.js framework with React.js and TypeScript to build the frontend and then hosted the backend on Firebase, which allowed us to fully integrate its NoSQL database and file storage services. For generating latent-consistent images to pass on, we used the Stable Diffusion API from getimg.ai, and tuned the hyper-parameters depending on the game mode for the most fitting experiences.

Challenges we ran into

We ran into issues with CORS (Cross-Origin Resource Sharing) while trying to fetch images stored in firebase. With some research and following this thread, we realized that we can change the configurations of our firebase project through the associated Google Cloud terminal.

Additionally, we considered the possibility that multiple users could access the website at the same time. This meant that users could caption the same image in the same game. In order to counter this, we made our game multi-threaded, meaning if two users captioned the same image, we would create two separate threads for new users to play from. This decision inspired us to structure our firebase data as a tree with each user-submitted caption pointing to its parent node.

Accomplishments that we're proud of

There were a few milestones that we were especially proud of. First of all, we managed to tune the hyper parameter in such a way to provide latent-consistent images. For example, tuning "strength" allowed us to manipulate the degree of influence the new prompt has on the current picture. Second, we provide different game modes with different hyper parameters for users to interact with. Users try to keep the pictures as similar as possible in survival mode, the creative mode allows space for imagination, and the last mode allows space for complete chaos. Last but not least, we have a gallery with a "story-line" flow for users to clearly visualize and interpret how the chain of images influence one another.

What we learned

We learned about how images are viewed by diffusion models through thorough experimentation of its limits, during which we also learned about hyper-parameters that govern latent consistency across images. We also learned how to integrate and leverage different technology from a variety of sources to create versatile applications.

What's next for TelePhoto

We would love for users to have the ability to share their galleries with other users, or even just to download it for them to post on social media and share with their friends!
Currently, all of our games are public and open to all users. A future step could be to create private games for individuals to play exclusively. This could be a fun exercise for English learners in classroom settings.
We could experiment with new (and maybe limited) game modes for users periodically!