Inspiration

My first time in HCMC, and both shocked and in awe of the traffic and how some drivers are able to weave through it. The skilled motorbike driver became the inspiration for this project.

Games have always been better with a crowd. But usually the crowd can only choose from options someone already built. I wanted to see what happens when you let the audience type anything and AI makes it real inside the game.

What if someone types "dragon" and a fire-breathing parade snake blocks two lanes? What if someone types "angry grandma" and she's sitting on a plastic stool throwing slippers at you?

What it does

Player plays the game on their laptop. Swipe or arrow keys to dodge, tap to boost.

Audience opens a URL on their phone, types anything, and it appears on screen as a real obstacle with AI-generated pixel art, sound effects, and behaviors. "Dragon" gets a 7-segment chain that breathes fire. "Angry grandma" sits on a plastic stool throwing slippers at you. "Traffic jam" blocks all three lanes with one gap to squeeze through. The whole thing takes about 3 seconds from suggestion to obstacle.

After game over, the game shows a "Traffic Report" your stats, your rating (from "Tourist" to "Huyền Thoại Giao Thông")

Challenges

How do we balance fidelity in generation with real-time performance?

(1) A custom JSONL template for the receiving AI to turn prompt given by audience into an appropriate sprite with its own characteristics, to also include sound effect, game behaviour, and text descriptors

(2) We have various "levels" of generation - (i) cached templates for very common items, that are mixed and matched, (ii) 20 polygon constructions for sprites, to hit ~3s generation requirement, and (iii), 50 polygon construction for higher quality sprites in scenarios where delay is tolerable

Accomplishments that we're proud of

Someone types "angry grandma" and 3 seconds later, a grandma on a plastic stool is throwing pixel-art slippers at the player. A very fun, amusing and inspirational thing to realise - that this project was possible. Fully generated user content - sprites, sounds, descriptors. Optimising around the performance challenges was incredibly fulfilling within these 36 hours.

What's next for Saigon Rush

Streaming - Twitch/YouTube chat as the audience input. Every streamer's chat becomes the game input. Events - Team buildings, product launches, conferences. QR scan to playing in 10 seconds, no setup needed. Templates for other cities and events - Bangkok, Jakarta, Mumbai. Same engine, different chaos. The prompts just need local flavor.

Fundamentally becomes a powerful interactive brand engagement tool.

Note

Devpost doesn't accept Loom in video submission, so posted link under "Try it out"!

Built With

Share this project:

Updates