Export
Tamy Boubekeur
· 11 previous affiliations
- Télécom ParisSince 2013
- Full Professor · Since 2013
- Institut Mines-TélécomSince 2008
- Full Professor (on leave) · Since 2013
- Associate Professor · 2008–2013
- École PolytechniqueSince 2017
- Full Professor · Since 2017
- Adobe Systems (United States)Since 2019
- Senior Director & Senior Principal Scientist, Adobe Research · Since 2022
- Director & Principal Scientist, Adobe Research · 2021–2022
- Director & Principal Scientist, 3DI Research · 2019–2021
- Allegorithmic2016–2019
- Chief Scientist · 2016–2019
- Telecom Paris2008–2013
- Associate Professor · 2008–2013
- Université Paris-Sud2011–2012
- HDR, Computer Science · 2011–2012
- Technische Universität Berlin2007–2008
- Postdoc · 2007–2008
- University of British Columbia2004–2007
- Computer Science · 2004–2007
- PhD student · 2004–2007
- Université de Bordeaux2002–2004
- MS student · 2002–2004
Research Overview
Tamy Boubekeur focuses on 3D computer graphics, with specific expertise in geometry processing, real-time rendering, and the synthesis of procedural materials. His work addresses the computational challenges of modeling and visualizing complex 3D scenes, moving from fundamental data structures for point-based surfaces to modern generative models that utilize large multimodal models for material authoring. Currently a Senior Director and Senior Principal Research Scientist at Adobe Research, he leads Adobe Research France while maintaining an academic role as a professor at Ecole Polytechnique.
Research Trajectory
The early phase of Boubekeur’s research, spanning approximately 2004 to 2008, was primarily concerned with the representation and manipulation of 3D geometry, particularly point clouds and subdivision surfaces. During this period, point-based graphics emerged as a viable alternative to traditional polygonal meshes for handling high-resolution scanned data. His work focused on bridging the gap between raw point data and hardware-accelerated rendering pipelines. A significant contribution during this time was the development of techniques for local surface reconstruction and visualization, which allowed for the efficient display of complex point-sampled objects by generating overlapping 2-manifold patches. He also addressed the scalability of interactive shape editing, developing methods to make freeform deformation (FFD) independent of the underlying model's complexity through sampling-reconstruction frameworks.
Between 2008 and 2016, while serving as a professor at Telecom Paris, Boubekeur’s focus expanded into real-time rendering and global illumination. This period marked a transition from purely geometric concerns to the simulation of light transport in complex environments. He investigated stochastic approaches to global illumination, aiming to achieve realistic diffuse interactions in real-time, particularly for large-scale CAD models. This work sought to balance the high computational cost of physics-based light simulation with the interactivity requirements of professional design applications. His research also touched upon digital human modeling and interactive 3D navigation, reflecting a broader interest in the end-to-end pipeline of virtual content creation and interaction.
From 2017 onwards, Boubekeur’s trajectory has been increasingly defined by his leadership roles in industry, first as Chief Scientist at Allegorithmic and subsequently at Adobe. His research in this phase has centered on the intersection of traditional computer graphics and machine learning. A major theme has been the optimization of Monte Carlo rendering through advanced sampling structures, such as double hierarchies that approximate light fields to improve importance sampling. Most recently, his work has shifted toward the automation of material authoring. He has explored generative models for procedural material graphs, addressing the difficulty of creating complex, resolution-independent textures. This current direction involves leveraging large multimodal models (LMMs) to synthesize "visual programs"—node-based graphs that define material appearance—by allowing models to perceive and iterate upon intermediate visual results.
Current Directions
Boubekeur’s current research at Adobe Research France focuses on the evolution of procedural modeling and the integration of generative AI into the graphics pipeline. A primary objective is the development of tools that lower the barrier to entry for creating high-quality 3D content. This includes the synthesis of procedural materials that are both parametric and resolution-independent, moving away from manual node-based authoring toward automated or semi-automated systems.
His recent work explores the capacity of Large Language Models and Multimodal Models to understand and generate the complex, unstructured graphs used in professional material design. By treating material authoring as a program synthesis problem, his team aims to enable the creation of infinitely detailed surfaces through natural language or multimodal prompts. This research also continues to address fundamental problems in rendering efficiency, ensuring that these complex materials can be visualized accurately and performantly in real-time environments.
Representative Work
In Reconstruction locale et visualisation de nuages de points par surfaces de subdivision, Boubekeur addressed the limitations of existing hardware architectures in handling point-sampled geometry. The paper proposed a technique for the efficient visualization of point clouds by performing local surface reconstruction via triangulation. By generating an aggregate of overlapping 2-manifolds, the method allowed for high-quality rendering of complex objects directly within the hardware rendering pipeline, providing a more economical way to store and transmit complex 3D models compared to traditional polygonal descriptions.
The problem of scalability in interactive 3D shape editing was addressed in SIMOD: Making Freeform Deformation Size-Insensitive. As the first author, Boubekeur introduced a streaming system based on a sampling-reconstruction approach. The system utilized an efficient out-of-core adaptive simplification algorithm to generate a simplified version of a model for interactive deformation. The deformations applied to the simplified proxy were then propagated back to the high-resolution original, ensuring that interactivity was maintained even when the complexity of the 3D model exceeded the workstation's immediate processing capacity.
Boubekeur also contributed to the field of digital human modeling, as seen in Guest Editors' Introduction: Digital Human Faces: From Creation to Emotion. This work surveyed the state of the art in the creation and animation of digital faces, covering applications ranging from basic synthesis to the representation of complex emotions in embodied conversational agents. This editorial contribution highlighted the multidisciplinary nature of his work, connecting geometric modeling with the requirements of human-computer interaction and virtual character design.
In the domain of realistic synthesis, A stochastic and scalable approach for real-time global illumination presented a novel method for addressing the high computational cost of light transport simulation. As the senior author, Boubekeur guided the development of a forward rendering technique focused on unshadowed diffuse interactions for the first indirect light bounce. Inspired by light cuts, the approach provided a scalable solution for real-time global illumination in large-scale models, such as those found in complex CAD scenarios, where traditional path-tracing methods were historically too slow for interactive use.
The efficiency of Monte Carlo rendering was further improved in Double hierarchies for efficient sampling in Monte Carlo rendering. This work proposed a new representation of the light field tailored for importance sampling. By hierarchically subdividing the product space of spatial positions and directions using a double-tree data structure, the method enabled efficient queries and progressive refinement. This framework allowed for the characterization of light transport using local models, significantly improving the convergence of Monte Carlo simulations by focusing computational effort on the most significant light contributions.
More recently, Boubekeur has focused on the intersection of AI and procedural content. In MultiMat: Multimodal Program Synthesis for Procedural Materials using Large Multimodal Models, he explored the use of Large Multimodal Models for the synthesis of procedural material graphs. The research addressed the specific challenge of "visual programs"—node-based graphs that define surfaces like rusted metal or woven fabric. Unlike standard code generation, material synthesis requires the model to understand the visual output of intermediate nodes. The paper introduced a multimodal approach that allows models to "see" the results of their work, enabling the automated creation of high-quality, parametric materials that were previously the sole domain of expert technical artists.
Beyond these core contributions, Boubekeur has participated in research regarding hybrid deformation structures, such as in the work on Real-Time Deformation with Coupled Cages and Skeletons, which sought to unify two popular but often disconnected approaches to character rigging. He has also been involved in the development of generative models for procedural materials, such as MatFormer: a generative model for procedural materials, which addressed the unstructured and long-range connections inherent in material operation graphs. These collaborative efforts underscore his role in advancing the state of the art in both the underlying mathematics of geometry and the practical tools used in modern 3D production.
Profile updated
| Title | Venue | Year |
|---|---|---|
| MultiMat: Multimodal Program Synthesis for Procedural Materials using Large Multimodal Models | ICLR | 2026 |
| Lipschitz Pruning: Hierarchical Simplification of Primitive-Based SDFs | Computer Graphics Forum | 2025 |
| Triangle Rejection Sampling for Density‐Equipped Meshes on GPU | Computer | 2025 |
| Φeat: Physically-Grounded Feature Representation | arXiv | 2025 |
| Evaluating and Sampling Glinty NDFs in Constant Time | ACM Trans. Graphics | 2025 |
| N-BVH: Neural ray queries with bounding volume hierarchies | Siggraph (conference Paper Track) | 2024 |
| MatUp: Repurposing Image Upsamplers for SVBRDFs | Computer Graphics Forum | 2024 |
| A Survey on Cage-based Deformation of 3D Models | Computer Graphics Forum | 2024 |
| Variational Shape Reconstruction via Quadric Error Metrics | Siggraph (conference Paper Track) | 2023 |
| MesoGen: Designing Procedural On-Surface Stranded Mesostructures | Siggraph (conference Paper Track) | 2023 |
Loading new papers…
Other researchers whose papers cluster into the same research directions — a proxy for overlapping technical interests and shared intellectual neighborhoods.
See something wrong? Report a mistake