Krea AI has released an open-source version of its image generation model, designed to produce visuals with a distinct aesthetic and avoid the common look of AI-generated content. In their post authors Sangwu Lee and Erwann Millon detailed the philosophy and process behind their new model, FLUX.1 Krea. The model was developed in collaboration with Black Forest Labs.
The company’s primary goal was to address what it calls the “AI look”, a collection of traits like overly blurry backgrounds, waxy skin textures, and uninspired compositions that make many AI images feel generic. Krea states that in the pursuit of technical benchmarks, the stylistic diversity and creative feel of earlier models have been lost. The team argues that common metrics for measuring model performance and aesthetics are often misaligned with user preferences and can even introduce biases into models.
To create a model with a more opinionated style, Krea adopted a two-stage training approach. The first stage, pre-training, focused on giving the model a broad understanding of the visual world to maximize diversity. The company argues that this stage should even include “bad” data, so the model can learn what to avoid when guided by negative prompts.
The second stage, post-training, is where the model’s aesthetic is shaped. Krea compares this process to Michelangelo’s belief that a sculpture already exists within a block of marble and only needs the excess material chiseled away. For this, the team started with a “raw” base model from Black Forest Labs, which was not yet heavily refined.
This raw model then underwent Supervised Finetuning (SFT) using a small, carefully curated dataset of fewer than one million images that matched Krea’s specific aesthetic standards. Following this, the model was further refined using Reinforcement Learning from Human Feedback (RLHF), where it was optimized based on internal preference data.
Krea argues that training a model on broad, “global” user preferences is suboptimal, as it tends to create results that satisfy no one completely. Instead, the company deliberately took a focused approach to align the model with a specific artistic direction. This method, they claim, provides users with a strong stylistic base that requires less extensive prompting to achieve high-quality results.
The model, named FLUX.1 Krea [dev], is now publicly available for download on Hugging Face. You can also try it for free.