Master's student — AI, Algorithms & Decision — Sorbonne University
Frédéric
Li Combeau
I build and study generative models — the kind that don't just render a plausible image, but hold some usable notion of the scene behind it. I am currently looking ahead to a PhD in generative AI — models that hold some real understanding of the world they're generating, representation learning and/or world models. Feel free to reach out if you have an open position.
Now Research intern at the PRS Group, ETH Zürich, working with Google on text-to-panoramic 3D Gaussian Splatting via MeanFlow distillation and model stitching.
01Research
-
Text-to-Panoramic 3D Gaussian Splatting via MeanFlow Distillation and Model Stitching
PRS Group, ETH Zürich — in collaboration with Google
Supervised by Postdoc. Narnhofer, Prof. Schindler, Prof. Tombari
Master's thesis project. A paper is planned for submission after the thesis is done.
-
Limitations of Test-Time Adaptation (TTA) methods in Online Continual Learning
CVM Lab, University of Tokyo
Supervised by Postdoc Michel, Prof. Yamasaki
Proposed a new, more realistic evaluation setup for combined continual-learning / test-time-adaptation deployment. A paper is planned for submission soon.
02Academic Projects
-
M2 project
Style Decomposition and Content Preservation for Artistic Style Transfer
Reproduction of the original paper plus additional experiments.
Original paper ↗ -
M2 project
Multimodal Visual Personalization of Generative Models via Individual Preference Learning
Added a multimodal (eye-tracking) signal and a new pipeline to improve image personalization, with a full evaluation of the resulting model.
Paper website ↗ -
M2 project
highest grade
Tactical Optimism and Pessimism in RL
Analysis of the original paper plus additional experiments.
Original paper ↗ -
M1 semester project
highest grade
AFU: Actor-Free Critic Updates
Implementation and analysis of “AFU: Actor-Free critic Updates in off-policy RL for continuous control.” Found that AFU outperforms other algorithms in offline-to-online transition, but not in an off-policy uniform-sampling setting.
Github ↗ -
Bachelor's project
Smart Home
A small smart house built on Raspberry Pi and Arduino: MobileNet for person detection, a Firebase database, a mobile app and a PC interface.
03Background
Master's degree, AI2D (AI, Algorithms, Decision)
Sorbonne University
Deep Learning · Advanced Deep Learning (audited) · Multimodal AI · AI for Robotics (ranked 1st) · AI for Vision (extra course) · Advanced Optimization Methods and Applications · Human–AI Interaction
Extra course, Robotics
ENS Ulm
Bachelor's degree, Computer Science
Paris-Saclay University
Programming
Python, PyTorch, OpenCV, NumPy, SciPy, Pandas, scikit-learn, Seaborn, Git/GitHub
Languages
- French — native
- English — professional (TOEIC C1)
Other
- French driving licence
- Civil Security citizen skills certificate
- Certificate of Musical Studies (11+ years of guitar and composition)
04Side Projects
Smaller things, built outside of coursework, mostly to understand something by hand.
-
Deep Learning / TensorFlow - 2023
DCGAN for face generation
The faces are not very nice but I only trained for 20 epochs.
Github ↗ -
Deep Learning / TensorFlow - 2023
Object detection
Full implementation of YOLO from scratch for object detection.
Github ↗ -
Deep Learning / TensorFlow - 2023
Emotion detection
Implementation of a deep learning model for detecting emotions in facial expressions.
Github ↗ -
Deep Learning / TensorFlow - 2023
Malaria detection
Implementation of multiple deep learning models and data augmentation techniques for detecting malaria in blood smears.
Github ↗ -
Computer Vision / OpenCV / MediaPipe - 2022
Detection and tracking
A set of computer vision experiments: face detection, hand tracking, pose estimation, and holistic (combined face + hands + pose) tracking.
Github ↗
05Off the clock
My hobbies and interests outside of work.
- Ongoing Composer, guitarist and pianist.
- 2024 Took part in Prologin, the French programming competition.
- 2018 Competed in the France Chess Championship at the regional level.
- Middle school A few charity sporting events for Action contre la Faim.
- Always History and hiking, usually not at the same time.
Here is a small selection of the games and anime I enjoyed the most.