Eric-Tuan Lê
Research Scientist · Meta AI
I am a Research Scientist at Meta AI developing multimodal reasoning capabilities in the Llama vision–language models. My work centers on post-training (SFT → RLHF/RLAIF) to enhance step-by-step reasoning, factual consistency, and instruction-following across text–image tasks. I have contributed to Llama 3 & 4 through structured-image alignment (charts, tables, forms, and UIs) and large-scale synthetic QA generation pipelines to expand data coverage and reasoning diversity.
Before joining the Llama team, I worked in GenAI 3D at Meta, conducting research on 3D and 4D generative modeling -- training diffusion and flow-matching models over signed distance fields (SDFs) and extending score distillation sampling (SDS) to dynamic scene generation.
Previously, I earned a PhD in Computer Science from University College London (Smart Geometry Processing Group) co-advised by Iasonas Kokkinos and Niloy J. Mitra, specializing in deep learning for 3D vision and graphics. My prior research experience includes Meta AI, Adobe Research (ICCV 2021 + patent), and Snap Inc. (CVPR 2024 + patent). I serve as a reviewer for CVPR, ICCV, ECCV, and NeurIPS.
Research
MeshPose: Unifying DensePose and 3D Body Mesh Reconstruction
CVPR 2024
Method unifying DensePose and Human Mesh Reconstruction, achieving significantly higher 2D DensePose accuracy while maintaining on-par 3D reconstruction quality and real-time performance for AR applications.
StyleMorph: Disentangled 3D-Aware Image Synthesis with a 3D Morphable StyleGAN
ICLR 2023
3D-aware generative model bridging morphable models and 3D-aware GANs; generates shapes through a morphable template with dense correspondence, enabling disentangled control over shape, pose, and appearance.
Work Experience
Meta AI
Research Scientist — London, UK (Mar 2024 – Present)
Working across two research efforts within GenAI and Llama organizations:
Research Intern (GenAI 3D) — London, UK (Jun 2023 – Dec 2023)
Extended Score Distillation Sampling (SDS) to 4D (3D+time) generation, improving temporal coherence and geometric consistency for dynamic scene modeling. Built pipelines for time-consistent mesh and texture extraction from volumetric reconstructions.
Research Scientist — London, UK (Mar 2024 – Present)
Working across two research efforts within GenAI and Llama organizations:
- Llama Post-Training (Multimodal: Text + Image): Contributed to advancing multimodal reasoning models through supervised fine-tuning (SFT) and reinforcement learning (RLHF/RLAIF), improving step-by-step reasoning, factual consistency, and instruction-following. Also worked on structured-image alignment (charts, tables, forms, UIs) and synthetic QA generation pipelines to broaden multimodal data coverage and task diversity.
- GenAI 3D: Trained and evaluated flow-matching models for 3D shape editing and enhancement by learning time-conditioned velocity fields over signed distance functions (SDFs). Improved geometric fidelity, edit controllability, and stability through few-step ODE sampling, classifier-free guidance, and ablations on path parameterizations.
Research Intern (GenAI 3D) — London, UK (Jun 2023 – Dec 2023)
Extended Score Distillation Sampling (SDS) to 4D (3D+time) generation, improving temporal coherence and geometric consistency for dynamic scene modeling. Built pipelines for time-consistent mesh and texture extraction from volumetric reconstructions.
Snap Inc.
Research Intern — London, UK (Jun 2022 – Dec 2022)
Developed MeshPose (CVPR 2024 + patent), a real-time framework unifying DensePose and 3D body mesh recovery for augmented-reality applications. Designed geometry-aware losses aligning 2D DensePose UVs with mesh supervision, achieving over 100 FPS inference and state-of-the-art 2D reprojection accuracy while maintaining competitive 3D performance.
Research Intern — London, UK (Jun 2022 – Dec 2022)
Developed MeshPose (CVPR 2024 + patent), a real-time framework unifying DensePose and 3D body mesh recovery for augmented-reality applications. Designed geometry-aware losses aligning 2D DensePose UVs with mesh supervision, achieving over 100 FPS inference and state-of-the-art 2D reprojection accuracy while maintaining competitive 3D performance.
Adobe Research
Research Intern — San Jose, USA (Jun 2020 – Nov 2020)
Designed CPFN (ICCV 2021 + patent), a cascaded primitive-fitting network for robust geometric reconstruction from high-resolution 3D point clouds. Introduced an adaptive patch-sampling strategy that jointly leverages global and local detections for state-of-the-art accuracy.
Research Intern — San Jose, USA (Jun 2020 – Nov 2020)
Designed CPFN (ICCV 2021 + patent), a cascaded primitive-fitting network for robust geometric reconstruction from high-resolution 3D point clouds. Introduced an adaptive patch-sampling strategy that jointly leverages global and local detections for state-of-the-art accuracy.
Patents
2023
Unifying DensePose and 3D Body Mesh Reconstruction
(US20240303937A1) — filed by Snap Inc.
Real-time framework unifying 2D dense correspondence (DensePose) and 3D body mesh estimation for AR, enabling efficient and accurate human reconstruction on mobile devices.
Real-time framework unifying 2D dense correspondence (DensePose) and 3D body mesh estimation for AR, enabling efficient and accurate human reconstruction on mobile devices.
2021
Fitting 3D Primitives to a High-Resolution Point Cloud
(US20220292765A1) — filed by Adobe Inc.
Cascaded neural architecture for detecting and fitting geometric primitives in large-scale 3D scans, providing fast, interpretable, and high-fidelity 3D reconstruction.
Cascaded neural architecture for detecting and fitting geometric primitives in large-scale 3D scans, providing fast, interpretable, and high-fidelity 3D reconstruction.
Teaching & Service
Reviewer
CVPR, ICCV, ECCV, NeurIPS (2021–2025)
Teaching Assistant
Computer Vision, Machine Learning, Computer Graphics, 3D Geometry Processing, and Information Retrieval (UCL, 2018–2021)
Supervision
MSc projects (CentraleSupélec, 2017): NLP and sentiment-based financial risk prediction
Contact
Location
London, United Kingdom