Training-Free 3D Inpainting via Initial Noise Optimization
A new method called InpaintSLat enables controllable 3D inpainting without training, using initial noise optimization within a structured 3D latent diffusion framework. The approach addresses instability in tasks requiring strict alignment with existing context while synthesizing new structures. It updates the initial noise via backpropagation approximation based on a rectified flow model, with spectral parameterization for robust optimization. Experiments show consistent improvements in fidelity. The paper is available on arXiv.
Key facts
- InpaintSLat is a training-free approach for controllable 3D inpainting.
- It uses initial noise optimization in a structured 3D latent diffusion framework.
- The method updates initial noise via backpropagation approximation grounded in rectified flow model.
- Spectral parameterization is used for robust and efficient optimization.
- Experiments demonstrate consistent improvement in 3D inpainting fidelity.
- The paper is published on arXiv with ID 2605.00664.
- The approach addresses stability issues in inpainting and editing tasks.
- Geometric structure is established early in diffusion and is sensitive to initial noise.
Entities
Institutions
- arXiv