2024 - Present · Darmstadt
PhD Researcher · Multimodal AI Lab, TU Darmstadt
PhD research under the supervision of Prof. Dr. Marcus Rohrbach and Prof. Dr. Anna Rohrbach, focused on reliable and controllable multimodal generative models.
PhD Candidate · Multimodal AI · Darmstadt, Germany
Jonas Henry
Grebe
I study how generative models can be used and controlled, and where their failures create practical risks.
Outside work, I like imagined worlds, creative tools, travelling, and swimming.
SEOUL / FIELD 01
ICML 2026 · Seoul
Concept erasure and robustness
GEM develops a geometric erasure method, while Erased but Not Forgotten examines how erasure can fail under backdoors.
01 / 03
Recent updates / rolling log
News.
A compact record of awards, conference acceptances, talks, releases, and other milestones.
Experience
Experience.
Selected roles and academic background.
May 2023 - Dec. 2023 · Berlin
Applied Scientist · Amazon Development Center
Ran experiments to improve machine-learning models for future deployments and built tooling that streamlined the deployment process.
04 earlier records Show full experience Collapse experience
Sep. 2022 - Mar. 2023 · Darmstadt
Student Research Assistant · Fraunhofer IGD
Developed diffusion-based synthetic face-data generation for a master's thesis, later published at ICCV 2023.
May 2022 - Aug. 2022 · Edinburgh
Applied Scientist Intern · Amazon Development Centre
Built a multimodal classification proof of concept for identifying catalog items that were safe to promote in ads.
Apr. 2020 - Mar. 2023
M.Sc. Computer Science · TU Darmstadt
Grade 1.28. Thesis on identity synthesis via latent diffusion models for face recognition, graded 1.0.
Oct. 2016 - Jan. 2020
B.Sc. Computer Science · TU Darmstadt
Grade 1.34 with a focus on machine learning. Thesis on anomaly-based face search, graded 1.0.
Selected publications / 2023 / 2026
Research.
I work on controlling generative models, examining how they fail or can be misused, and using them for visual creation. Browse by method or theme, or see the full publication list on Google Scholar ↗
* Equal contribution.
Personal Projects.
Open-source creative workbench / Exploratory worldbuilding
Fogwood Studio.
ALPHA / 0.1An open-source, local-first environment for creating with generative models.
Fogwood Studio brings image generation, editing, drawing, notes, references, and visual history into one deliberate workspace. Model choice, prompts, and creative history remain visible throughout the process.
View on GitHub ↗
Apache 2.0
