Inked
A four-player free-for-all platform fighter where drawing a shape on a canvas casts the spell, with GPU shape recognition in Unity.
Media
Project Overview
Inked is a four-player free-for-all platform fighter that six of us built over twelve weeks for the Game Projects course at Howest DAE. It is set in a hand-crafted origami world. You cast a spell by drawing its shape: the controller paints your lines onto a canvas texture in real time, and when you stop, the finished stroke is classified on the GPU and the matching spell is equipped.
I was one of two programmers and owned the recognition pipeline. I trained a Convolutional Neural Network (CNN) in PyTorch on the Google Quick, Draw! dataset to tell six hand-drawn shapes apart, then ran it in real time through Unity Sentis. The rest of my work was gameplay systems, VFX and shaders.
My Contribution
Draw-to-cast: Players draw on the canvas with the controller, and the lines paint onto a texture as they move.
Shape recognition: When the drawing stops, the texture goes to a Convolutional Neural Network (CNN) I trained in PyTorch on the Google Quick Draw! dataset. It picks the closest of six shapes and equips that spell.
Real-time inference: Unity Sentis runs the model on the GPU, and an MPSC queue manager for async inference.
Generalist role: On a team of six I built gameplay systems, integrating VFX, shaders, and I owned the recognition pipeline.
SYS_SPECS
- Language:
- C#, Python, HLSL
- Technology:
- Unity3D, Pytorch
- Platforms:
- Windows

