RFDM: Residual Flow Diffusion Model for Efficient Causal Video Editing
CVPR 2026
RFDM is a causal video editing model that edits variable-length videos efficiently by exploiting temporal redundancy through residual flow diffusion.
[PDF]
I am a Researcher working on efficient generative models and world models.
My research focuses on diffusion models, image and video generation, 3D understanding and geometry-aware representations, and post-training.
I am interested in building controllable and efficient generative models with applications in spatial intelligence, robotics, and interactive world simulation.
I completed my PhD at University College London, supervised by Niloy J. Mitra and Noam Aigerman.
My studies were funded by UKRI Centre for Doctoral Training in Foundational Artificial Intelligence.
Before that, I completed my MSc and BSc at Politecnico di Milano.
CVPR 2026
RFDM is a causal video editing model that edits variable-length videos efficiently by exploiting temporal redundancy through residual flow diffusion.
[PDF]
ICML 2026
NanoFLUX compresses a 17B FLUX text-to-image model into a 2.4B model for high-quality, low-latency generation on mobile devices.
[PDF]
CVPR Findings 2026
FraQAT introduces fractional-bit quantization aware training to progressively reduce model precision while preserving generative model quality.
[PDF]
arXiv 2025
Explicit Conditioning reduces the need for classifier-free guidance in image editing, improving diffusion inference speed while preserving quality.
[PDF]
ICCV 2025
EDiT introduces linear compressed attention for diffusion transformers, improving text-to-image efficiency while retaining high-quality generation.
[PDF]
ICML 2025
Multi-Task Upcycling extends pre-trained text-to-image diffusion models to multiple image-to-image tasks using expert FFNs and dynamic routing.
[PDF]
PhD thesis, University College London, 2024
This thesis discusses map-based neural representations as effective alternative to meshes in geometry processing tasks.
[PDF]
Eurographics 2024
Through pre-trained ViT(DinoV2) we extract (visual) semantic correspondences to optimize inter-surface mapping.
[PDF]
[Webpage]
[Code]
[Video]
CVPR 2022
Neural surface representation disentangling coarse geometry from fine details. Thanks to CNNs inductive bias, the description is compact and allows editing....
[PDF]
[Webpage]
[Code]
CVPR 2021
Neural representation of surface maps, allowing optimization of differentiable quantities in inter-surface maps.
[PDF]
[Webpage]
[Code]
ICRA 2019
Fast scene point cloud simplification leveraging semantic information.
[PDF]
[Webpage]
IAS-15 2018
Next Best View prediction based on photogrammetry principles.
[PDF]
[Webpage]
[Code]
Future Internet 2018
Novel player represenation in Physically-Interactive RoboGames.
[PDF]
[Webpage]
[Code]