Krea drops open weights to cure the "AI look"

FLUX.1 Krea [dev] is an open-weights 12B parameter text-to-image model built by Krea in partnership with Black Forest Labs. It is architecturally compatible with the FLUX.1-dev pipeline, meaning you can drop it directly into your existing diffusion setups.

The entire point of this release is to cure the dreaded "AI look." Instead of spitting out images with waxy plastic skin, washed-out palettes, and aggressively blurred backgrounds, it is tuned to generate authentic, photorealistic shots with distinct creative character.

What it does

If you have generated photos with open diffusion models lately, you know the telltale signs: skin looks like polished rubber, lighting feels synthetic, and every portrait has the exact same depth of field. FLUX.1 Krea [dev] tackles these artifacts head-on.

It produces portraits with believable skin texture, moody cinematic lighting, and naturally balanced compositions. In human preference assessments, the team notes it outranks existing open-weights models and performs on par with proprietary commercial models like FLUX1.1 [pro], all while retaining FLUX's strong text-rendering and prompt-following capabilities.

How it was built

Most foundation models train on broad aesthetic scorers like LAION-Aesthetics, which heavily skew outputs toward over-softened, overly bright, and generic visuals. Krea took the opposite route by taking an intentionally opinionated approach.

They started with flux-dev-raw, an unbaked 12B parameter guidance-distilled diffusion transformer from Black Forest Labs that had diverse world knowledge without any baked-in visual bias.

For post-training, Krea ditched massive data scraping and used fewer than 1M meticulously curated images. They applied supervised fine-tuning with high-quality synthetic samples, followed by multiple rounds of Reinforcement Learning from Human Feedback using an internal preference optimization technique called TPO.

Their philosophy was simple: subjective beauty is personal. Blending everyone's aesthetic preferences results in bland, plastic mediocrity, while focused, curated data can deliberately push a model toward high-end realism.

Where to run it

The model weights are live on Black Forest Labs' Hugging Face repository, ready for local inference in FLUX-compatible environments. Commercial licenses are handled through the BFL Licensing Portal.

If you do not have local compute, cloud API endpoints are already available via partners including FAL, Replicate, Runware, DataCrunch, and TogetherAI.

Why it matters

For designers, concept artists, and photographers tired of wrestling with endless negative prompts and LoRAs just to strip away the plastic AI sheen, this checkpoint is an immediate upgrade. It delivers genuinely usable, cinematic results out of the box.

It also proves that post-training quality trumps raw quantity. By showing that fewer than one million focused samples can rival top closed-source image generators, Krea and BFL offer a clear blueprint for building specialized, high-taste open models.