DeepSOM: Generative Manipulation for 3D Texture Synthesis

Problem and context

Detailed 3D objects are difficult to manipulate directly, and large paired 3D datasets are scarce. DeepSOM asks whether a simple proxy can become an intuitive control surface for a more complex form.

My role

I developed this academic project with Amit Zoran at the Hebrew University of Jerusalem, working across the research question, training pipeline, 3D representation, and public demonstration.

Constraints

The method starts from a single source object and primitive proxy. Training examples are produced through controlled free-form deformations rather than a large external dataset.

What I built

The model learns to transfer proxy deformations to a detailed 3D object, allowing small edits to generate varied spatial textures and forms.

Key decisions

The simple proxy is the interaction model; the detailed object remains the output. Synthetic paired supervision turns a limited source example into a trainable manipulation problem.

Evidence

The public project page contains the academic abstract, demonstration video, authorship, institutional attribution, and rendered texture examples.

Results

The demonstration shows the proxy and complex object moving together, making the research question understandable without requiring access to the training system.

Limitations and status

This is presented as a completed thesis or academic project. The legacy public page contains incomplete sections, so this case study links only to artifacts that currently resolve and does not claim a paper publication.