While he waited, Leo explored the project’s raw data. The Subfloor was set entirely in a single, infinite basement. The logs showed SkeletonCrew had added a “shadow actor” – a prop that wasn’t a prop, but a negative space. A cutout in the digital world. In the script notes, SkeletonCrew had written: “The monster isn’t added. It’s subtracted.”
The (Vertex-Voting Variational 3D Morphable Model) represents a significant advancement in statistical shape modeling. Unlike traditional Morphable Models (3DMMs) that rely heavily on pre-established dense correspondences (often achieved through expensive Non-Rigid ICP or manual landmarking), v3DMM utilizes a probabilistic framework. It treats shape correspondence as a latent variable within a Variational Autoencoder (VAE) framework, allowing the model to learn a deformable template directly from unorganized, unregistered 3D point clouds or meshes. This approach mitigates the "chicken-and-egg" problem where good registration is required to build the model, but a good model is required to achieve registration. While he waited, Leo explored the project’s raw data
“He lives between the polygons.”
The screen went black. A single line of green text appeared in the center, rendered in the old Comic Sans MS font: A cutout in the digital world
v3DMM can be trained on datasets where shapes are not pre-aligned or registered. The network learns to identify "semantically" similar regions (e.g., the tip of a nose on a face, or a specific anatomical landmark on an organ) purely through the statistical regularities of the data. but didn't close the program.
Leo tried to pause. The spacebar did nothing. He tried to open the editing menu. It was grayed out. The escape key cycled the camera from Buster to first-person, but didn't close the program.