Background
Esser worked with Björn Ommer at Heidelberg and LMU Munich, where 'Taming Transformers for High-Resolution Image Synthesis' (VQGAN) came out of the group and became a foundation of the text-to-image era. He joined Stability AI, was a lead author on Stable Diffusion 3's rectified-flow scaling paper, and co-founded Black Forest Labs in 2024.
He is the least publicly visible of the three founders. His most quotable public appearance remains the August 2024 a16z podcast the founders did together around the FLUX.1 launch.
What they run now
Core model research across the FLUX line. His public framing of the product problem is about controllability and prompting: he has said they trained the model to respond well to varied prompting techniques, and that the human remains the one who decides which of a hundred generations is right for a project. Inside BFL that maps to prompt-following rubrics, editing fidelity and the sample-selection loop — all of which are rubric-and-rater problems.
What they buy
Inference. Prompt-following and editing-fidelity rubrics, expert pairwise preference data for image editing (the Kontext line), and human-judged sample selection for SFT sets. No named vendor. FLUX.1 Kontext shipped with a crowd-sourced KontextBench, which shows willingness to use external raters at least once.
Career
- 2024 – presentCo-founder, Black Forest Labs
- 2021 – 2024Researcher (Stable Diffusion, SD3), Stability AI
- 2017 – 2021PhD researcher, Heidelberg University / LMU Munich (Ommer group) — VQGAN / Taming Transformers
- c. 2017–2021Heidelberg University, PhD, computer vision (Ommer group)
On the record
Says taste and curation of the sample thread stay with the person: you can make a hundred samples fast, but which one serves your project is your call. source
Notes they deliberately trained the model to react well to a range of prompting techniques. source