Sangwu Lee

Sangwu Lee is head of AI at Krea, where he develops image-generation models that give artists and designers greater creative control. His work on FLUX.1 Krea and Krea 2 challenges a central tradeoff in generative media: models optimized for reliable results often converge on the same predictable aesthetic.

Earlier, at the University of Rochester, Lee worked on computer vision and multimodal machine learning. His coauthored research on multimodal transformers incorporated visual and acoustic signals into pretrained language models; subsequent work explored humor recognition and video-based assessment of Parkinson’s disease.

At Krea, he helped develop FLUX.1 Krea with Black Forest Labs and coauthored the July 2025 open-weight release, detailing how aesthetic post-training could adapt a general-purpose image model for creative work. He subsequently led research on Krea’s own image foundation-model family as first-listed author of the Krea 2 technical report. The June 2026 release included openly available model variants and official inference code.

• Creative exploration over aesthetic conformity. Lee prioritizes fast generation, stylistic range, and unconventional compositions because creative studios often need to discover what they want through iteration. He avoids training heavily on synthetic images that transfer another generator’s recognizable house style.

• Dataset curation as creative direction. He preserves visually unconventional material that standard quality filters might discard, combines optical character recognition with vision-language captions, and uses hashes and compact classifiers to remove duplicates and unwanted imagery. When captions repeatedly omit details such as a painting’s white background, he treats the resulting hidden association as a training-data problem.

• Sparse autoencoders for visual filtering. Lee applies sparse autoencoders to vision-model representations, producing unsupervised visual features that can identify watermarks, signatures, borders, and other recurring artifacts at scale.

• Visual world knowledge with practical specialization. He checks coverage against prominent Wikipedia concepts, then develops capabilities for illustration, photography, and graphic design through progressively higher-resolution training, supervised fine-tuning, preference optimization, and reinforcement learning. His account of Krea 2’s research also identifies specialist-model distillation, richer visual conditioning, and simpler image-generation architectures as promising next steps.