Blog #2: From Pixels to Personalization – The Tech Behind the Avatar

In my previous post I talked about the “why” of my virtual fitting room. Now, let’s get to the “how.” That is, how do I take two basic photos from a smartphone and turn them into a dynamic, anatomically realistic 3D mesh?

I have blended RealityScan and MetaHuman concepts to morph my MakeHuman base models to avoid the "too smooth" look of generic software. My pipeline uses a powerful open-source engine named Godot to host the 3D environment. The magic begins with Otsu’s thresholding, which produces a binarized image and extracts a clean silhouette. The system then automatically finds key landmarks to measure based on existing anthropometric ratios; for example, the waist should be three-eighths of a person’s height.

The real challenge is not only to create a static model but also to let it walk. I can rig the skeleton with Mixamo so I can try out dresses in dynamic poses like “reaching” or "stepping." This is where Enhanced Position-Based Dynamics (PBD) comes in. PBD, in contrast to traditional simulations, enables me to observe real-time "draglines," diagonal wrinkles indicating a garment is too tight, thus allowing me to perform a subjective fit analysis that numerical data alone might miss.

I'm currently struggling to implement high-fidelity cloth physics in Godot directly. Is this technical deep dive into morphing and landmark detection making sense, or do I need to explain more about the specifics of the “3/8ths rule” math?