New AI method generates consistent 3D multi-view scenes from rough hand-drawn sketches, solving a previously unsolved computer vision challenge.

Researchers tackle the first-ever problem of creating geometrically consistent multi-view 3D scenes from single freehand sketches, which contain spatial distortions that conflict with 3D interpretation

The team created a curated dataset of approximately 9,000 sketch-to-multiview samples using an automated generation and filtering pipeline to address the lack of training data

Introduced Parallel Camera-Aware Attention Adapters (CA3) that inject geometric reasoning into the model to handle distorted 2D sketch input and maintain cross-view consistency

This approach fills a gap between existing methods: multi-view generators that require photographs or text, and sketch-to-3D methods that need multiple views or expensive per-scene optimization