Laurent Lucas
Professor
LICIIS - LRC DIGIT
University of Reims
IUT de Reims
Chemin des Rouliers
CS 30012
51687 Reims Cedex 2
FRANCE
laurent.lucas (AT) univ-reims.fr
+33(0)3 2691 8452
2025 Projects
|
Whatever project you choose, please consider the following deliverables:
Projet 1 - ANARI: A 3D rendering API standard ANARI is a new 3D rendering API and emerging Khronos standard designed to let visualization applications easily use advanced rendering techniques across different hardware and engines. Traditionally, such applications relied on custom-built renderers to meet user demands for performance and quality, but maintaining these has become increasingly costly and complex as rendering technology evolves. ANARI solves this problem by offering a high-level, visualization-focused API that hides low-level hardware and algorithm details, allowing applications to efficiently access various ANARI implementations and take advantage of the latest rendering capabilities. In this project, you consider implementing and experimenting with the ANARI API to better understand its architecture, flexibility, and performance across rendering backends. You are expected to reproduce the core examples and, where possible, extend the framework. Potential extensions include benchmarking ANARI against native rendering APIs (e.g., OptiX), integrating ANARI into your own visualization tools, exploring hybrid rendering pipelines combining rasterization and ray tracing, or contributing new extensions or adapters to support emerging rendering engines or novel visualization use cases. Keywords: programming volume rendering, ANARI Some references:
Other links: Projet 2 - Multivariate volumetric rendering Multi-channel volume rendering has become essential for interpreting modern scientific datasets that encode several measurements per voxel, yet existing visualization pipelines still struggle to provide physically meaningful blending and intuitive exploration tools. Recent advances in Monte Carlo (MC) volume rendering offer high flexibility and robustness, but their use in multi-channel visualization remains limited. A promising direction is the extension of physically based MC techniques—such as multi-density Woodcock tracking—to handle multiple channels without relying on arbitrary color blending rules, enabling more faithful and interpretable visualizations. Within this master project, you will investigate and prototype methods for multi-channel volume rendering that unify physically based blending, high-dimensional transfer-function design, and efficient viewpoint or interaction strategies. The goal is to develop a principled, interactive rendering framework capable of handling complex multi-channel data while maintaining visual quality and real-time usability. Potential extensions include exploring alternative MC sampling strategies, integrating dimensionality reduction for multi-channel transfer functions, or designing adaptive accumulation and denoising schemes tailored to multi-channel datasets.. Keywords: multi-channel volume rendering Some references:
Other links: Projet 3 - LLM-based function transfer Traditional volume visualization (VolVis) methods such as Direct Volume Rendering (DVR) rely on transfer functions (TFs) to reveal structures in volumetric data [Ljung2016], but these approaches remain difficult to design and interact with, especially for non-experts. To overcome these limitations, NLI4VolVis [Ai2025] enables natural language interaction for exploring, querying, and editing volumetric scenes, while IntuiTF [Wang2025] leverages Multimodal Large Language Models (MLLMs) to make TF optimization more intuitive and perceptually aligned. Combining semantic understanding, language-based interaction, and intelligent TF exploration, these frameworks bridge the gap between user intent and volumetric visualization. In this project, you will implement and extend these approaches to develop a unified, language-driven volume visualization system. Possible extensions include integrating reinforcement learning for adaptive TF refinement, using diffusion models for visual evaluation, or extending the system to dynamic and multimodal volumetric datasets. Keywords: transfer function, DVR, multimodal LLM Some references:
Other links: Projet 4 - AI-assisted temporal volumetric rendering Time-series forecasting plays a crucial role across domains such as healthcare, finance, and simulation, where improved prediction accuracy can significantly enhance decision-making and efficiency. While deep learning (DL) models have shown strong performance in forecasting tasks, they typically require extensive training and validation before being applied to new datasets. To overcome this limitation, [Das2024] introduces TimesFM, a foundation model trained on a large corpus of real-world time points. TimesFM achieves strong zero-shot performance across diverse datasets and temporal resolutions, approaching the results of state-of-the-art supervised models while requiring no additional training. In this project, you will study how language models dedicated to time series processing can be considered for managing temporal sequences of 3D volumes. Keywords: time series, analysis, multimodal LLM Some references:
Other links: Projet 5 - Out-fo-core on hybrid architecture The need of visualizing and/or processing large volume data has become common today in different fields such as biomedical or even entertainment. Navigating inside such volumes in real-time involves designing efficient out-of-core data management algorithms to address entire massive datasets from high-performance computing devices such as current GPUs. [Sar2020] introduce a complete out-of-core pipeline from disk to GPU to access very large volumes exceeding GPU or CPU memory in real-time. They base their work on modern methods already known in this field, such as the design of output-sensitive algorithm, the on demand-paging and data streaming, the bricking and multi-resolution epresentation, the use of a brick pool as a cache on the GPU, the virtual addresses translation and more. However, they propose a system that allows access to multi-modal volumes from any type of end-user application with a GPU interface connected to our out-of-core pipeline. This project will involve setting up accurate (GPU timing, performance measurements, etc.) and reproducible benchmarks for Out-Of-Core systems on hybrid architectures. Preliminary tests and usage scenarios will be carried out, and remote visualization on HPC systems will be implemented using an Out-Of-Core solution. An experimental approach will then be used to evaluate the application's performance on NVIDIA's Grace-Hopper hybrid architectures, compared to the hardware and software capabilities available on the ROMEO platform. In addition to hybrid architectures and depending on progress, other solutions may also be considered in order to extend the data exchange pipeline itself, for example by precaching bricks on the RAM. Some references:
Other links:
Projet 6 - LLM-based voice user interface Large Language Models (LLMs) are fundamentally changing interaction: LLM-driven Voice User Interfaces (VUIs) ofer more natural interaction, better handle ambiguity and the unpredictability of spoken language, and support complex multi-turn conversations, outperforming traditional rule-based systems. Their advanced capabilities are driving a paradigm shift from using systems as tools for the execution of commands with a precise expectation of the outcome to experiencing them as collaborative partners, capable of interpreting intentions and generating context-driven content (system as a partner). This idea evolved in the feld of HCI, where researchers have explored how such systems and humans can jointly solve problems as active participants. In this project, you will explore how these ideas can be put into practice by building a prototype integrating an LLM and enabling interaction through both VUI and GUI modalities. In this prototype, spoken natural language commands are interpreted by the LLM and translated into high-level intentions that drive the behavior of the graphical user interface, allowing users to control, configure, and manipulate the system using voice in a flexible and context-aware manner. For example, a user can verbally request to open a specific view, adjust parameters, navigate within the interface, or trigger complex actions that would traditionally require multiple manual interactions. This multimodal approach illustrates how a voice-driven interface can act as an intelligent mediator between the user and the GUI, supporting more natural, efficient, and collaborative human–system interaction. Keywords: Speech-to-text, LLM Some references:
Other links: |