SemanticSlider3D lets you tweak 3D objects with sliders
New method lets designers edit 3D objects in real-time without training models
Ru Wang and three co-authors present SemanticSlider3D, a training-free method for continuous semantic editing of 3D objects. Given a user-specified attribute, it builds a semantic editing direction in the latent space of a state-of-the-art 3D generation model, yielding a diverse and coherent spectrum of 3D variations. In a technical validation with 50 3D object-attribute pairs, all five human assessors preferred it over a baseline combining a 2D slider with an image-to-3D model across variation range, consistency, 3D object quality, and attribute disentanglement. An exploratory study with six participants found that SemanticSlider
- Training-free continuous semantic editing of 3D objects using sliders, developed by researchers from four universities.
- Validated on 50 object-attribute pairs with all assessors preferring it over a 2D slider + image-to-3D baseline.
- Enables real-time, fine-grained adjustments in latent space without model retraining, improving prototyping workflows.
Why It Matters
SemanticSlider3D streamlines 3D content creation, reducing reliance on manual modeling or expensive training for precise edits.