Alessio Mazzucchelli

I am a PhD student at the Institut de Robòtica i Informàtica Industrial (IRI), CSIC-UPC in Barcelona, working on 3D Computer Vision and 3D Gaussian Splatting. I was supervised by Prof. Francesc Moreno-Noguer until mid-2024, when he joined Amazon; I am now supervised by Adrian Penate-Sanchez and Jordi Sanchez-Riera.

Currently, I am a researcher at Fondazione Bruno Kessler (FBK) in Trento, part of the Deep Visual Learning (DVL) team led by Elisa Ricci. Previously, I worked as a research engineer at Arquimea, and I received my degree from Politecnico di Milano.

I'm interested in everything 3D Gaussian Splatting, mainly editing and segmentation.

Email  /  Google Scholar  /  Github  /  LinkedIn

• 2026: BEA-GS is accepted as a Highlight to CVPR 2026!
• 2026: VIRGi is accepted to IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)!
• 2026: G1NDiff is accepted to Computers & Graphics!
• 2024: IReNe is accepted to CVPR 2024!

BEA-GS introduces two new losses that shape the underlying 3D geometry of Gaussian Splatting scenes, producing near-perfect boundaries for object-level extraction without relying on radiance supervision alone.

VIRGi enables rapid, photorealistic recoloring of 3D Gaussian Splatting scenes from a single edited image, separating diffuse and view-dependent color components to preserve specular effects while propagating edits in seconds.

G1NDiff combines GAN and diffusion training on 2D datasets, using reprojection conditioning to enable fast, consistent 3D scene editing.

IReNe enables near real-time color editing of NeRF scenes from a single edited training image, fine-tuning only the last network layer's diffuse-appearance neurons to achieve 5x–500x speedups over prior work.

Note: IReNe, VIRGi, G1NDiff, and BEA-GS were all developed at Arquimea (my employer at the time), which has not released official code for these works. If you're interested in building an unofficial implementation of any of them, feel free to get in touch — happy to help.

A comparative study of Fourier transform-based and CycleGAN-based domain adaptation techniques for weed segmentation under distribution shift.