About This Project
This is a demo of a diffusion model that was trained on about 600 samples of real human mouse movement data. (from my friends and me collecting data, using the data collection page) The model is a 1D U-Net that refines (yes, refines) a straight line from x1,y1 to x2,y2 into a more human-like trajectory.
You can notice a few characteristics of human mistakes in the generated paths, like jitters and overshooting. But, it becomes an issue because literally almost every generated path has a banana-like curve to it. This is likey due to a small dataset, and a lack of diversity in movement styles.
One thing -- The model's banana-curve tendency can be somewhat mitigated by turning off clamping + smoothening, but then the model will just straight up overshoot, which isn't great either.
