Know3D lets users control the hidden back side of 3D objects with text prompts

Know3D lets users specify the unseen back side of a 3D object with text prompts during single-image 3D generation.

The system inserts an image generator between a multimodal language model and a 3D generator to translate text into spatial guidance.

Using intermediate internal states from the image model produced better semantic and geometric 3D backs than final images or extracted features.

Know3D scored highest on HY3D-Bench for semantic match and back-side geometry, but results depend on correct prompt interpretation by the language model.