The Invisible Backbone: 3D Representations for Intuitive Content Creation in the Era of GenAI
Abstract
With the rapid emergence of high-fidelity video generation models, the role of explicit 3D modeling is facing an existential question: if pixels can generate pixels, do we still need geometry? In this talk, I argue that while traditional manual modeling may be fading for non-experts, 3D representations are more critical than ever, not as a final output, but as the invisible backbone that guarantees consistency and control in generative workflows. I explore this thesis through two recent works from Adobe Research. First, I discuss Project Turn Style, a pipeline that uses 3D reconstruction as a transient proxy to enable one-shot rotation of 2D images. Second, I present SplatPainter, a framework for interactively editing 3D Gaussian Splats. By leveraging a feedforward architecture combined with Test-Time Training (TTT), SplatPainter allows users to propagate 2D edits, such as local texture painting or global relighting, instantly across a 3D asset without the destructive artifacts of diffusion or the latency of optimization.
