I am a research intern at Disney Research|Studios and hold a PhD from INRIA Paris under the supervision of Raoul de Charette. During my doctorate, I collaborated extensively and explored research at the intersection of computer vision and graphics. My work focuses on material estimation from images and controllable image generation. I have a background in physics and engineering sciences, later transitioning to computer science, and hold an MSc from Telecom SudParis after an exchange program at Linköping University and Telecom Paris.
This thesis explores material understanding and dense multi-task learning from images, presenting methods for extracting and editing physical surface properties, and multi-task models leveraging task correlations and diffusion priors trained on synthetic data.
A zero-shot multi-task dense prediction method that repurposes a diffusion model for latent regression across multiple tasks simultaneously. Trained on multiple synthetic datasets each annotated for only a disjoint subset of tasks.
MatSwap enables photorealistic transfer of a given material onto a target surface in an image, handling complex interactions of appearance, geometry, and lighting without requiring manual input or UV mapping.
A method to transfer material transformations across scenes. Building on disentangled NeRF representations, our approach learns to map Bidirectional Reflectance Distribution Functions (BRDF) from pairs of scenes observed in varying conditions.
Material Palette extracts a palette of PBR materials (albedo, normals, and roughness) from a single real-world image using diffusion priors.
This work proposes an architecture to perform multi-task learning by combining information from different tasks using correlation-guided and self-attention mechanisms.
End of course computer graphics project, coded in C and GLSL.
Proof of concept project for a node-based editor to create systems from a visual interface.