AI-generated imagery often suffers from a lack of distinctiveness — many enterprise integrations lead to visuals that feel generic, monotonous, and branded as "AI slop." Krea, a San Francisco startup founded in 2022 by Víctor Perez and Diego Rodriguez Prado, aims to change that narrative. On June 23, 2026, the company open-sourced the weights of its new Krea 2 model in two variants: Krea 2 Raw and Krea 2 Turbo. Both are available on Hugging Face under a custom license with specific usage terms.
Krea 2 Raw and Turbo: two checkpoints for different creative needs
Krea 2 Raw is an undistilled base checkpoint taken from the mid-training stage of the larger Krea 2 Medium development cycle. It lacks post-training alignment, RLHF, or aesthetic distillation, making it a blank canvas ideal for training custom LoRAs or fine-tunes without stylistic bias. Krea 2 Turbo, on the other hand, is a distilled variant optimized for speed: it generates native 2K resolution images in about 2 seconds on consumer-grade hardware, thanks to Trajectory Distribution Matching and an 8-step inference pipeline. It ranks among the fastest AI image generators, surpassed only by specialized models like FLUX.1 schnell (0.5 seconds) and Z-Image Turbo (1.8 seconds).
Sponsored Protocol
At the core is a 12-billion parameter Diffusion Transformer built from scratch. Krea adopts a single-stream transformer block with shared attention and MLP layers for text and image tokens, SwiGLU MLP with 4x expansion, Grouped-Query Attention, and gated sigmoid attention to stabilize training. Timestep conditioning uses a lightweight per-block bias term, cutting modulation parameters by 20–30%. Positional encoding uses 3D Axial Rotary Position Embedding (RoPE).
"Train on Raw, generate with Turbo" workflow for creators
Krea promotes a clear workflow: train LoRAs on the Raw checkpoint and then directly apply them to Turbo for fast generation. This preserves creative freedom while leveraging speed. The company itself has developed an in-house collection of custom LoRAs following this philosophy. Additionally, Krea 2 includes an advanced style transfer system: users can upload reference images, isolate aesthetic components, and adjust style strength via generative sliders. A LLM Prompt Expander, refined with Generalized Deep Q-Network Preference Optimization, enriches short prompts while preserving intent.
Sponsored Protocol
Custom license mandates content moderation for all hosts
The Krea 2 Community License permits free commercial use for individuals and small businesses with up to 50 seats. Beyond that, an enterprise agreement is required. However, unlike traditional open-source licenses (MIT, Apache 2.0), this license imposes strict downstream behavioral guardrails: any entity hosting the model must implement active input/output classifiers to prevent the generation of illegal materials, non-consensual intimate imagery (NCII), child sexual abuse material (CSAM), and defamatory content. Failure to do so constitutes a breach of contract, giving Krea the right to revoke access. The official Krea API remains a separate paid service.
Sponsored Protocol
An alternative to rigid closed-source APIs
Krea 2 Raw is viewed as a creative alternative to heavily aligned APIs from closed-source models. Its lack of aesthetic alignment allows artists to "break the model if they want to," fostering genuine artistic exploration. With this release, Krea transitions from a multi-model aggregator to a model provider in its own right. Partnerships with Superside (80% of their AI generation through Krea) and Henning Larsen for EU AI Act-compliant design tools underscore its enterprise maturity.
This open-weight release marks a turning point for enterprise AI image generation, balancing speed, customization, and control. For related reading, check our articles on Rate Limiting for APIs and the Tata Electronics cyberattack. For broader context, see the Wikipedia page on Diffusion Transformers.
