About
Manuel Brack is an Applied Research Scientist at Adobe Firefly in San Jose, USA, with an adjunct research affiliation at hessian.AI in Darmstadt, Germany. His academic foundation includes a Ph.D. from TU Darmstadt's Machine Learning Lab where he specialized in generative AI systems. He co-founded the Occiglot research collective focused on open-source LLM development.
Brack's research centers on large-scale generative models at the intersection of natural language processing and computer vision. His work addresses critical challenges in monosemanticity measurement, multilingual data curation, AI safety frameworks, and text-to-image generation ethics. He has pioneered techniques for bias mitigation in multilingual systems and developed novel architectures for memory-efficient language modeling.
His publications reveal significant trends toward neuro-symbolic integration (DeiSAM), privacy-preserving mechanisms (CLIP identity attacks), and community-driven open science (Occiglot/Community OSCAR). The research consistently bridges theoretical innovation with practical deployment considerations for generative systems.
- Best Runner-Up Paper Award RBFM Workshop at NeurIPS 2024 (LlavaGuard)
- Best Paper Award at DPFM 2024 (Homoglyphs research)
Brack actively contributes to open-source initiatives including Community OSCAR (345+ TiB multilingual dataset) and Occiglot (open LLM development). His work emphasizes ethical deployment through frameworks like Fair Diffusion for bias attenuation and Safe Latent Diffusion for content moderation, demonstrating commitment to responsible AI advancement.
Find Manuel Brack elsewhere
Related Searches
You Might Also Like
Sebastian FaustTechnical University of Darmstadt · Professor- OOliver FalckTechnical University of Darmstadt · Professor
Aryaman ReddiTechnical University of Darmstadt · Researcher- UUwe RüppelTechnical University of Darmstadt · Professor
- HHelmut RainerTechnical University of Darmstadt · Professor
Simon PoppingaTechnical University of Darmstadt · Adjunct Professor