Skip to main content
Back

Black Forest Labs and Krea AI's FLUX.1 Krea Fights AI's Generic Image Problem

Black Forest Labs and Krea released a 12B open-weights text-to-image model tuned to escape the oversaturated, waxy AI look.

  • Black Forest Labs and Krea AI released FLUX.1 Krea [dev], a 12B open-weights text-to-image model.
  • Distilled version of Krea 1, targeting photorealism and avoiding the oversaturated AI look.
  • Drop-in replacement for FLUX.1 [dev], works in ComfyUI and diffusers on day one.
  • Trained from a special flux-dev-raw base using SFT plus a preference optimization variant called TPO.
  • Weights are gated non-commercial; hosted APIs via FAL, Replicate, Runware, Together, DataCrunch.
  • Claimed on par with closed FLUX1.1 [pro] in human preference for aesthetics.

Text-to-image models keep getting technically better while producing images that look increasingly the same: blurred backgrounds, waxy skin, blown-out highlights, and boringly symmetric compositions. FLUX.1 Krea [dev], a collaboration between Black Forest Labs and Krea AI, attacks that problem directly. The 12-billion-parameter open-weights model trades benchmark chasing for opinionated aesthetics and drops in as a replacement anywhere FLUX.1 [dev] already runs.

What makes it different

FLUX.1 Krea [dev] is the open version of Krea's closed Krea-1 model, post-trained on a base checkpoint that Black Forest Labs provided specifically for this purpose. The team calls it "opinionated" meaning it makes deliberate aesthetic choices rather than averaging across user preferences. Despite those idiosyncrasies, BFL and Krea claim it outperforms previous open text-to-image models and matches closed solutions like FLUX1.1 [pro] in human preference evaluations. It is guidance-distilled and fully compatible with the existing FLUX.1 [dev] ecosystem: LoRAs, ControlNets, and ComfyUI workflows all carry over without modification.

Why the AI look keeps happening

Krea's technical writeup frames the sameness of modern generated images as a Goodhart's law problem. Older aesthetic scorers like LAION-Aesthetics, PickScore, ImageReward, and HPSv2 are mostly fine-tuned CLIP variants operating on 224×224 crops. As generators improved, those scorers stopped keeping up, and training pipelines that filter data through them inherit their taste: bright images, blurry backgrounds, soft textures, and a strong subject bias.