Original title: Rethinking 3D Noise: Learning 3D-Aware Video Priors via Optimization-Free Morphological Perturbations
Onat Şahin, Mohammad Altillawi, George Eskandar, Carlos Carbone +1 · September 2026
Recreating 3D scenes from only a few pictures often causes messy visual glitches. Recent tools try to clean up these glitches, but training them usually takes expensive, slow 3D rebuilds for every scene. Researchers found a faster way to train these fixers without those slow steps. They treated 3D Gaussian Splats—tiny 3D building blocks that act like 3D pixels—by directly altering them. They scrambled their shapes by scaling, rotating, and pruning them to instantly create realistic flaws that remain consistent across different camera angles. Tests showed this taught video models much stronger shapes and geometry. It cut depth errors by 12.5% compared to top previous fixers. It also improved robot task success by up to 8.0% on three of four chores.
Imagine you're putting together a jigsaw puzzle but only have a few scattered pieces to start with. Normally, you'd struggle with gaps or mismatches because you can't see how everything fits. Here, the method gently nudges each piece—changing its shape or angle slightly—while making sure they still lock together perfectly, so the full picture emerges without needing every piece upfront.
This could help robots better understand and interact with 3D environments, making manipulation tasks more successful and reliable. It reduces errors in 3D models used for navigation and control.
The abstract notes evaluations were conducted on a lightweight video diffusion sandbox and scaled to a 14B-parameter video model via ControlNet, so it does not establish generalizability to other 3D representations or broader applications beyond these specific setups.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Rethinking 3D Noise: Learning 3D-Aware Video Priors via Optimization-Free Morphological Perturbations — Onat Şahin, Mohammad Altillawi, George Eskandar et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.