i'm an ai researcher working on pre-training of generative models at google deepmind in the genmedia team.
advising. i hosted 8 interns in my first 2.5 years at google. a full list is in my cv.
natively multimodal models.
- i worked on large-scale pretraining and new capabilities for gemini omni, an any-to-any modality multimodal model with frontier video generation capabilities. gemini omni flash was released at google i/o 2026.
image generative models.
- i am the lead author of dreambooth, an early personalization method for text-to-image diffusion models. it was one of the five awarded best papers at cvpr 2023.
- i continued this direction with methods for fast personalization like hyperdreambooth (and suti).
- i worked on camera coach, announced at made by google 2025, on a feature called ideations that lets a user generate inspiring alternate views and compositions of a scene when taking photos on a google pixel phone.
- i worked on styledrop and ziplora, which are, respectively, the first method for style personalization and the first method for style/subject lora merging.
- i led magic insert, the first working style-aware drag-and-drop method for compositing subjects into images. it was a highlight at iccv 2025. i also worked on realfill, the first modern method for authentic image completion from reference images.
video generative models.
- i worked on recapture, which brings camera control to user-provided videos by re-rendering a single clip from new viewpoints.
- we made a non-local video editing method that lets you control motion, timing, and camera movement with point trajectories, called motionv2v.
generative games and world models.
- i worked on unbounded, a generative infinite game of character life simulation with open-ended mechanics generated in real time.
- we made multigen, one of the first multiplayer generative games. it allows for an arbitrary number of players, real-time online play, and custom designs of maps and levels. we had a demo where anyone could play online.
deepfakes and adversarial attacks.
- i worked on adversarial attacks against deepfake generation, which we called disrupting deepfakes and later worked on black-box attack extensions.
training and testing using simulation.
- earlier, i worked on simulating synthetic data for training and testing neural networks. the work started with learning to simulate and continued with several other works. my phd thesis is about this.
background. bachelors from ecole polytechnique (paris), masters in computer science from georgia tech, phd in computer science from boston university. i am bolivian and french, and i live in boston, usa (which is a wonderful place btw). i deeply admire the founding principles of the united states.
values and beliefs. i believe we are in the midst of a new industrial revolution with the advent of ai. i believe that humans can find solutions to the hardest problems and am optimistic about our ability to improve the world, persist in history and improve the condition of humanity. in interpersonal relationships i believe in kindness, honesty, and aligned incentives. i value working on things that are at the edge of knowledge and i enjoy learning.