Background
Millon came out of the open-source generative-art community (VQGAN+CLIP-era work is associated with his GitHub) into Krea as a founding engineer. He co-wrote the FLUX.1 Krea release in July 2025, describing the SFT-then-RLHF pipeline and the TPO preference-optimisation variant trained on rigorously filtered internal preference data.
In October 2025 he authored Krea Realtime 14B, a 14-billion-parameter real-time long-form video model distilled from Wan 2.1 14B, hitting 11fps in four inference steps on a single B200, with streaming prompt changes mid-generation. That post evaluates quality qualitatively — comparing checkpoints by eye — which is a candid admission of where formal evaluation is missing.
What they run now
Krea's model engineering: post-training and preference optimisation on the image side, and real-time video generation and distillation. The Krea Realtime work is the newest surface — long-form, streaming, interactive video — and it is exactly the kind of model whose quality regressions are invisible to automatic metrics and obvious to a trained eye.
What they buy
Inference, constrained by the same in-house stance as Sangwu Lee. The genuine gap visible in his own writing is the video side: Krea Realtime's quality claims rest on the authors' own eyes, with no rater panel and no rubric. A small, fast, expert panel scoring real-time video checkpoints on motion dynamism, high-frequency detail and temporal artifacts is a product-shaped hole he has publicly described.
Career
- c. 2023 – presentFounding Engineer / ML Researcher, Krea
On the record
Co-author of the claim that Krea's RLHF used high-quality internal preference data, rigorously filtered, from labellers who knew the model's weaknesses. source
Assesses real-time video checkpoints by observing motion dynamism and high-frequency detail rather than by formal human study. source
