Krea releases open weights for FLUX.1 Krea
Krea has released open weights for FLUX.1 Krea, a model built with Black Forest Labs that targets the “AI look” and drops into existing FLUX.1-dev workflows.
Krea announced on 31 July 2025 that it is releasing an open version of Krea 1, its first image model, which it trained with Black Forest Labs. The checkpoint, FLUX.1-Krea [dev], is a guidance-distilled model that is fully compatible with FLUX.1-dev. That means it can plug into the existing FLUX tooling. Krea says the checkpoint was distilled to match Krea 1's quality, with an emphasis on keeping its aesthetics and photorealism.
The post, by Sangwu Lee and Erwann Millon, presents the model as deliberately "opinionated." Its stated goal is to "make AI images that don't look AI," which means avoiding overly blurry backgrounds, waxy skin and dull composition. The authors argue that common benchmarks miss these problems. DPG, GenEval, T2I-CompBench and GenAI-Bench mostly measure prompt adherence. Popular aesthetic scorers such as LAION-Aesthetics, PickScore, ImageReward and HPSv2 are mostly CLIP fine-tunes that work on 224×224 images. Krea also says LAION Aesthetics, which is widely used to filter training data, is biased toward images of women, blurry backgrounds, soft textures and bright images.
The accompanying technical report describes Krea's training approach. Pre-training should aim for "mode coverage" and broad knowledge of the world, even if that means including "bad" data. Krea argues that negative prompts only work if the model has learned what undesirable outputs look like. Post-training then pushes the model toward the outputs Krea prefers, which the company calls "mode collapse."
The excerpt available ends before the post-training method is described in full. It also gives no benchmark results, licensing terms or download details.