Patrick Pérez

Patrick Pérez is a French computer scientist whose research spans probabilistic image analysis, motion estimation, large-scale image retrieval, photo editing, and autonomous-driving perception. He is co-author on the Poisson Image Editing paper (SIGGRAPH 2003), the exemplar-based image inpainting paper (IEEE TIP 2004), the VLAD paper (IEEE PAMI 2012), and the Moshi paper (2024). As of May 2026, he is Chief Executive Officer of Kyutai, the Paris-based nonprofit AI research lab he joined at its December 2023 launch alongside Hervé Jegou, Edouard Grave, and three other founding researchers.

At a glance

• Education: Engineer's degree, École Centrale Paris (Mathématiques Appliquées option), 1990; PhD in Signal Processing and Telecommunications, Université de Rennes 1 and Inria, 1993, advised by Fabrice Heitz. Postdoctoral researcher in Applied Mathematics at Brown University, 1993 to 1994.
• Current role: Chief Executive Officer at Kyutai since December 2023.
• Key contributions: Poisson Image Editing (SIGGRAPH 2003); Region filling and object removal by exemplar-based image inpainting (IEEE TIP 2004); the VLAD compact-image-representation paper (IEEE PAMI 2012); the ADVENT domain-adaptation paper (CVPR 2019); the Deep Reinforcement Learning for Autonomous Driving survey (IEEE T-ITS 2021); co-author on Moshi (2024).
• LinkedIn: patrick-pérez-253695146
• Personal site: ptrckprz.github.io
• Google Scholar: 8Cph5uQAAAAJ
• OpenReview: Patrick_Perez1

Origins

Pérez was born and raised in France. He entered École Centrale Paris (now part of CentraleSupélec) and graduated in 1990 with the engineer's degree in the Applied Mathematics option, recorded in the Légifrance decree publishing the 1990 ECP roster.

He carried out his doctoral research at the IRISA laboratory at Inria Rennes under Fabrice Heitz, defending in 1993 with a thesis on Markov random fields and multiresolution image analysis applied to motion ("Champs markoviens et analyse multirésolution de l'image"). The early line on stochastic image models shaped the probabilistic-vision orientation of the next two decades of his record. He spent 1993 to 1994 as a postdoctoral researcher in Applied Mathematics at Brown University.

Career

After his Brown postdoc, Pérez was hired in 1994 as a permanent researcher (chargé de recherche) at Inria Rennes, joining the Vista project-team led by Patrick Bouthemy on probabilistic models for moving-image analysis. The 1994-to-2000 period produced his early line on motion segmentation, Bayesian tracking, and stochastic image models.

In March 2000 Pérez moved to Microsoft Research Cambridge as a researcher in the computer-vision group led by Andrew Blake. The Cambridge period through February 2004 produced two of his most-cited papers. The Poisson Image Editing paper at SIGGRAPH 2003, with Michel Gangnet and Blake, introduced gradient-domain methods for seamless image cloning, texture transfer, and illumination editing, with over 4,000 citations and incorporation into mainstream image editors. The exemplar-based image inpainting paper at IEEE TIP in 2004, with Antonio Criminisi and Kentaro Toyama, became one of the most-cited classical inpainting methods.

In February 2004, Pérez returned to Inria Rennes as research director (directeur de recherche) within Vista and succeeded Bouthemy as team leader in 2007. The 2004-to-2009 period included work on stochastic tracking, video segmentation, and image retrieval, and the early collaborations with Hervé Jegou, Matthijs Douze, and Cordelia Schmid at Inria Grenoble that fed into the VLAD and product-quantization line.

In November 2009 he moved to Thomson Corporate Research (which became Technicolor, later InterDigital) in Rennes as senior researcher and Distinguished Scientist. The 2009-to-2018 period produced the Aggregating local image descriptors into compact codes paper at IEEE PAMI 2012 with Jegou, Florent Perronnin, Douze, Jorge Sánchez, and Schmid, formalizing the VLAD (Vector of Locally Aggregated Descriptors) line at over 6,000 citations.

In March 2018, Pérez was appointed VP of AI and Scientific Director of Valeo.ai, the Paris AI research lab founded in June 2017 by Valeo. He led the lab through 2023, building it into a leading European industrial-AI lab focused on autonomous-driving perception. The period produced the ADVENT paper at CVPR 2019 on adversarial entropy minimization for domain adaptation, and the Deep Reinforcement Learning for Autonomous Driving survey at IEEE T-ITS in 2021. Valeo.ai also formed the ASTRA joint Inria-Valeo research team.

In December 2023, Pérez left Valeo to join Kyutai at its launch as CEO, presenting alongside founding researchers Edouard Grave, Hervé Jegou, Laurent Mazaré, Neil Zeghidour, and Alexandre Défossez at Scaleway's ai-PULSE conference in Paris on November 17, 2023. Kyutai launched as a French nonprofit AI research lab funded by Iliad, CMA CGM, and Schmidt Futures with a €300 million ($330 million) initial commitment. Pérez is a co-author on the 2024 Moshi speech-text foundation model paper and presented the Moshi keynote on July 3, 2024.

Affiliations

• École Centrale Paris: Engineering student, Applied Mathematics option, graduated 1990.
• Université de Rennes 1 and Inria (IRISA): PhD candidate in Signal Processing and Telecommunications, advised by Fabrice Heitz, defended 1993.
• Brown University: Postdoctoral researcher in Applied Mathematics, 1993 to 1994.
• Inria Rennes (Vista project-team): Chargé de recherche, 1994 to 2000.
• Microsoft Research Cambridge: Researcher, March 2000 to February 2004.
• Inria Rennes (Vista project-team): Directeur de recherche, 2004 to 2009; team leader from 2007.
• Technicolor (formerly Thomson, later InterDigital): Senior researcher and Distinguished Scientist, 2009 to 2018.
• Valeo.ai: VP of AI and Scientific Director, 2018 to 2023.
• Kyutai: Chief Executive Officer, December 2023 to present.

Notable contributions

Pérez's record runs from probabilistic motion-analysis work at Inria, through the Microsoft Research Cambridge image-editing period, the VLAD image-retrieval line, the Valeo.ai autonomous-driving line, and the Kyutai speech-AI period. His Google Scholar profile lists approximately 51,700 citations and an h-index of 89 as of mid-2026, with nearly 300 co-authored publications and around 30 filed patents.

• Poisson Image Editing (SIGGRAPH 2003). With Michel Gangnet and Andrew Blake at Microsoft Research Cambridge; gradient-domain methods for seamless image cloning, texture transfer, and illumination editing, with over 4,000 citations.
• Region filling and object removal by exemplar-based image inpainting (IEEE TIP 2004). With Antonio Criminisi and Kentaro Toyama; over 4,700 citations on the exemplar-based inpainting approach.
• Aggregating local image descriptors into compact codes (IEEE PAMI 2012). With Jegou, Perronnin, Douze, Sánchez, and Schmid; over 6,000 citations on the VLAD compact-image-representation line.
• Color-based probabilistic tracking (ECCV 2002). With Carine Hue, Jaco Vermaak, and Gangnet; an early high-impact paper on particle-filter color tracking with over 2,500 citations.
• ADVENT: Adversarial Entropy Minimization for Domain Adaptation in Semantic Segmentation (CVPR 2019). With Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, and Mathieu Cord; over 2,000 citations on adversarial entropy minimization for unsupervised domain adaptation.
• Deep Reinforcement Learning for Autonomous Driving: A Survey (IEEE T-ITS 2021). With Ravi Kiran, Ibrahim Sobh, Victor Talpaert, Patrick Mannion, Ahmad Sallab, and Senthil Yogamani; over 3,500 citations on the deep-RL-for-autonomous-driving research literature.
• Moshi: a speech-text foundation model for real-time dialogue (2024). Co-author on the Kyutai full-duplex speech-text foundation model paper, with Alexandre Défossez, Laurent Mazaré, Manu Orsini, Amélie Royer, Hervé Jegou, Edouard Grave, and Neil Zeghidour.
• Editorial roles. Associate Editor of IEEE Transactions on Image Processing (2002 to 2006) and IEEE Transactions on Pattern Analysis and Machine Intelligence (2008 to 2012); Editorial Board member of the International Journal of Computer Vision (2010 to 2020). Area chair for ECCV (2008, 2010, 2012, 2014, 2020), ICCV (2011, 2013, 2019, 2023 industrial chair), and BMVC (2017).

Investments and boards

No public investor activity on record in AI, semiconductors, datacenters, software, or energy as of May 2026.

Network

Pérez's collaborations trace through four overlapping cohorts. The first is his Inria PhD lineage and early Inria Rennes period: Fabrice Heitz was his doctoral advisor, and Patrick Bouthemy led the Vista project-team he later inherited. The longest-running scientific collaboration through the Inria, Technicolor, and Valeo years was with Hervé Jegou, Matthijs Douze, and Cordelia Schmid on the Hamming-embedding, product-quantization, and VLAD image-retrieval lines.

The second is the Microsoft Research Cambridge cohort. The Poisson Image Editing and inpainting papers ran with Andrew Blake, Antonio Criminisi, Kentaro Toyama, and Michel Gangnet, with the probabilistic-tracking line adding Carine Hue and Jaco Vermaak.

The third is Valeo.ai, where Pérez built and led the Paris AI research team from 2018 to 2023. The autonomous-driving and domain-adaptation lines ran with Mathieu Cord (Sorbonne joint chair), Tuan-Hung Vu, Maxime Bucher, Himalaya Jain, and Senthil Yogamani.

The fourth is the Kyutai founding cohort: Edouard Grave, Laurent Mazaré, Neil Zeghidour, and Alexandre Défossez, alongside Jegou. The Moshi paper byline collects all six co-founders. The scientific committee includes Yann LeCun, Yejin Choi, and Bernhard Schölkopf.

Position in the field

As of May 2026, Pérez sits at the intersection of three research trajectories. The first is the classical-image-processing line through Poisson editing and image inpainting, which produced two of the most-cited image-editing methods of the 2000s. The second is the large-scale image-retrieval line through VLAD and color-based probabilistic tracking, which paralleled the Jegou-Schmid line and made the Inria-Rennes-and-Grenoble cohort one of the most-cited European computer-vision groups of the 2010s. The third is the autonomous-driving and applied-AI line through Valeo.ai and Kyutai, which moved his work from individual papers to leadership of European industrial and nonprofit AI labs.

His public-facing presence as a researcher remained low until Valeo.ai. As CEO of Kyutai he has presented the Moshi keynote on July 3, 2024 and at ai-PULSE 2024, and has appeared on French AI podcasts including Monde Numérique. The h-index of 89 and 51,700 citations, with editorial-board and area-chair service across ECCV, ICCV, BMVC, IEEE T-PAMI, IEEE T-IP, and IJCV from 2002 onward, place him among the most credentialed senior French computer-vision researchers of his generation.

Outlook

Open questions over the next 6 to 18 months:

• Kyutai senior recruitment. Whether the lab continues to attract senior FAIR Paris and other European-cohort talent at the cadence set in 2023 and 2024.
• Open-research and open-weights commitment. Whether Kyutai's positioning holds as the lab's compute footprint scales.
• Moshi successor and the speech-language stack. The trajectory of the speech-AI program, including any Moshi-2 or full-duplex multilingual successor.
• Hibiki and Helium iteration. Whether the Hibiki speech-to-speech and Helium compact-multilingual lines continue as recurring releases.
• European nonprofit AI positioning. Whether Kyutai's nonprofit posture (relative to Mistral AI and Aleph Alpha) attracts further strategic-investor commitments.
• Public-facing posture. Whether Pérez's communications cadence shifts toward broader European-AI-policy engagement.

Sources

• Patrick Pérez's personal site. Personal homepage with bio, research focus, and links to publications and recent activity.
• Patrick Pérez's Google Scholar profile. Citation metrics, h-index, and chronological publication record.
• Patrick Pérez's services and editorial-board page. Listing of editorial-board memberships, area-chair roles, and academic-committee service.
• Patrick Pérez's earlier homepage at Inria Rennes. Inria Vista-project-team page covering Pérez's PhD details, Brown postdoc, Microsoft Research Cambridge period, and 2007 Vista-team-leader appointment.
• Patrick Pérez on OpenReview. Career and education record.
• Légifrance decree confirming Pérez's 1990 École Centrale Paris graduation. The 1990 government decree publishing the École Centrale des Arts et Manufactures graduate roster.
• Patrick Pérez appointed Scientific Director of Valeo.ai. Valeo press release announcing the March 2018 Scientific Director appointment.
• Patrick Pérez, CEO de Kyutai: voyage au pays de l'IA. Inria Alumni interview covering Pérez's career trajectory and Kyutai launch.
• Kyutai team page. Lists Pérez with the Chief Executive Officer title.
• Valeo.ai team page. Lists Pérez in the Permanent member alumni section as former Scientific Director.
• Astra-vision team page. Lists Pérez in the alumni section of the Inria computer-vision research group descended from the Inria-Valeo joint ASTRA team.
• Poisson Image Editing. The 2003 SIGGRAPH paper introducing gradient-domain image-editing methods.
• Region filling and object removal by exemplar-based image inpainting. The 2004 IEEE TIP paper on exemplar-based inpainting.
• Aggregating local image descriptors into compact codes. The 2012 IEEE PAMI paper on the VLAD compact-image-representation line.
• ADVENT: Adversarial Entropy Minimization for Domain Adaptation in Semantic Segmentation. The 2019 CVPR paper on adversarial entropy minimization for unsupervised domain adaptation.
• Deep Reinforcement Learning for Autonomous Driving: A Survey. The 2021 IEEE T-ITS deep-RL-for-autonomous-driving survey.
• Moshi: a speech-text foundation model for real-time dialogue. The 2024 Kyutai full-duplex spoken-dialogue model paper.
• Unveiling of Moshi: the first voice-enabled AI openly accessible to all. Pérez's July 3, 2024 Moshi launch keynote on the official Kyutai YouTube channel.
• Kyutai launch announcement (TechCrunch). November 2023 coverage of the Kyutai launch and founding research team.
• Launch of Kyutai (Iliad press release). Iliad's launch announcement covering the November 2023 Kyutai unveiling.
• Photo: Kyutai Team page, used as the source portrait under press-use convention for the team-page photographs.