Refresh

Large language models are a new frontier in how we interact with and generate digital content, but they still face significant limitations when working with complex 3D objects. While models can partially reconstruct 3D geometry using tools such as Blender, maintaining geometric accuracy and consistency becomes increasingly difficult over longer generation processes, especially when working from image references, vertices, and other spatial data. Refresh is a Blender reconstruction frontend designed to make this process iterative and interactive. Users create local projects, capture or upload reference images, request an initial 3D reconstruction from a backend model, and inspect the resulting, models directly in the application. Rather than relying entirely on repeated model generations, users can then pinch-sculpt specific regions of the reconstructed object and send those targeted edits back to a refinement backend. The goal is to combine generative 3D reconstruction with direct human interaction, allowing users to correct and refine geometry where automated generation falls short. Browser-based hand tracking keeps the interaction layer responsive while the reconstruction and refinement processes are handled by the backend.