Publications

You can also find my articles on my Google Scholar profile.

Conference Papers

Qinco2: Vector Compression and Search with Improved Implicit Neural Codebooks

Théophane Vallaeys, Matthew Muckley, Jakob Verbeek, Matthijs Douze

Published in ICLR 2025, 2025

Improvements to QINCo for vector quantization with better encoding, fast approximate decoding, and optimized training, achieving state-of-the-art results in vector compression and billion-scale search.

Download Paper

Workshop Papers

Improved Baselines for Data-efficient Perceptual Augmentation of LLMs

Théophane Vallaeys, Mustafa Shukor, Matthieu Cord, Jakob Verbeek

Published in 1st Workshop on Green Fondation Models, ECCV, 2024

Experimental evaluation and improvements of data-efficient multi-modal adaptation of single-modality LLM and perceptual backbones.

Download Paper | Source code

arXiv Preprints

Beyond Language Modeling: An Exploration of Multimodal Pretraining

Shengbang Tong, David Fan, John Nguyen, Ellis Brown, Gaoyue Zhou, Shengyi Qian, Boyang Zheng, Théophane Vallaeys, Junlin Han, Rob Fergus, Naila Murray, Marjan Ghazvininejad, Mike Lewis, Nicolas Ballas, Amir Bar, Michael Rabbat, Jakob Verbeek, Luke Zettlemoyer, Koustuv Sinha, Yann LeCun, Saining Xie

Published in arXiv preprint, 2026

Empirical study of native multimodal pretraining using Transfusion framework, revealing key insights on visual representation, data synergy, world modeling, and MoE scaling.

Download Paper

VUGEN: Visual Understanding priors for GENeration

Xiangyi Chen, Théophane Vallaeys, Maha Elbayad, John Nguyen, Jakob Verbeek

Published in arXiv preprint, 2025

A framework that leverages VLM pretrained visual understanding priors for efficient and high-quality image generation, achieving superior performance while preserving understanding capabilities.

Download Paper

SSDD: Single-Step Diffusion Decoder for Efficient Image Tokenization

Théophane Vallaeys, Jakob Verbeek, Matthieu Cord

Published in arXiv preprint, 2025

A new pixel diffusion decoder architecture for image tokenization that achieves higher reconstruction quality and faster sampling than KL-VAE through distillation-based single-step decoding.

Download Paper