ViSculpt: Visual-Centric Agentic Geometry Editing

ViSculpt is a training-free, visual-centric multi-agent system for editing 3D meshes directly inside Blender. Instead of generating scripts or rebuilding geometry, it works through the native GUI: agents inspect the viewport, infer the mesh state, and execute localized changes with simulated user actions.

That matters because a lot of real production editing is about preserving what already works. In-place edits are a better fit for tasks like reshaping a part of an asset, adjusting proportions, or making a targeted correction without disturbing the rest of the model. The system is designed to stay visually grounded, which is important when the instruction is ambiguous and the untouched regions need to remain intact.

The work was submitted on 25 Aug 2026 and evaluated on a curated benchmark. The results are an early signal that agentic, GUI-driven editing can follow natural-language instructions and keep the overall identity of the input asset while making representative localized mesh edits.

For game teams, the practical angle is obvious: this is closer to how artists already work than script-first generation. It could eventually reduce friction in technical art and content iteration, especially for teams that want AI assistance without giving up control of the existing DCC workflow. The exact production limits are still being explored, but the direction is promising for Blender-centric pipelines.