Releasing open weights for FLUX.1 Krea

9 months ago

Release of FLUX.1 Krea [dev], an open version of Krea 1 for superior aesthetic control and image quality.
FLUX.1 Krea focuses on opinionated aesthetics to avoid the 'AI look' common in generated images.
Discussion on the limitations of current aesthetic evaluation models and their biases.
Training approach divided into pre-training (mode coverage and world understanding) and post-training (mode collapsing for aesthetics).
Use of a raw base model, flux-dev-raw, provided by Black Forest Labs for post-training.
Post-training pipeline includes Supervised Finetuning (SFT) and Reinforcement Learning from Human Feedback (RLHF).
Key findings: quality over quantity in data, the importance of an opinionated approach to aesthetics.
Future directions include personalization, aesthetics, and controllability research.
Acknowledgements to Black Forest Labs and internal teams for their contributions.

Hasty Briefsbeta