Krea, a startup building creative AI tools, has released Krea 2 (K2), a family of open-weights text-to-image foundation models that the company says ranks among the top 10 image generators on the Artificial Analysis leaderboard and second among models from independent labs. The release, detailed in a technical report published on June 23, 2026, ships with two model checkpoints and a permissive license that invites the community to fine-tune and build on top of it.
Unlike many recent image models that optimize for a single polished default output, Krea designed Krea 2 around the idea of creative exploration, exposing a broad visual distribution that users can navigate rather than forcing one narrow aesthetic. For more context on this story, see our ongoing breaking AI news.
Two Checkpoints for Different Needs
The Krea 2 release includes two distinct checkpoints. The first, K2 Raw, is an undistilled base model intended for fine-tuning and post-training research, giving developers a clean foundation to adapt. The second, K2 Turbo, is a guidance- and timestep-distilled model engineered for fast, few-step generation, the kind of quick iteration that creative workflows demand.
Offering both a research-friendly base and a speed-optimized variant reflects a pragmatic split. Researchers and tinkerers get the raw model to experiment with, while production users get a faster checkpoint that sacrifices little in the way of quality for the sake of speed.
A Deliberate Stance Against AI-Generated Training Data
One of the most striking claims in Krea's technical report concerns training data. The team states that it used no AI-generated images in its pretraining mix. While synthetic data and distillation have become popular shortcuts for acquiring model capabilities, Krea found that even a small proportion of AI-generated images introduced biases into the model's output distribution.
The reasoning is straightforward: synthetic images tend to be easier for a model to learn, which effectively imposes an artificial ceiling on quality. To enforce the policy, Krea built in-house classifiers specifically to filter AI-generated images out of its dataset, rather than relying on off-the-shelf filters.
The company also took a contrarian view on data quality. Rather than aggressively filtering for high aesthetic scores, which can remove deliberately blurry, grainy, or unconventional images that represent valid artistic choices, Krea argued for diversity and broad domain coverage. The result, it says, is a model with a wider expressive range than systems trained only on conventionally beautiful images.
Architecture and Training Pipeline
Krea 2 is built on a transformer architecture whose final design was chosen through extensive ablation studies documented in the report. After testing alternatives, the team settled on grouped-query attention (GQA) with gated sigmoid attention, a SwiGLU MLP, and Qwen 3 VL as its text encoder. Positional encoding uses 3D axial rotary embeddings, and the autoencoder combines the Qwen Image VAE and FLUX 2 AE.
Training followed a multi-stage curriculum. Pretraining progressed through 256px, 512px, and 1024px resolution stages, dedicating the bulk of compute to low-resolution work to build core capabilities efficiently before sharpening high-fidelity generation at higher resolutions. A midtraining stage then bridged the broad pretraining distribution and the high-quality supervised fine-tuning distribution using semantic clustering and retrieval-based strategies to preserve coverage of rare and long-tail concepts.
Making Generation Exploratory and Steerable
Krea identified a persistent gap between how models are trained and how users actually express intent. During training, models learn from rich, dense captions. At inference time, users often type short, ambiguous prompts, or gesture toward a mood or reference image rather than describing a scene precisely.
To close that gap, Krea 2 ships with two systems. The first is a prompt expander, trained through a two-stage supervised fine-tuning and reinforcement learning pipeline on top of open-source large language models, which maps simple prompts into richer visual directions without overwriting the user's intent. The second is a style-reference system that lets users inject the style or mood of one or more reference images with minimal content leakage, offering fine-grained control over style strength and weighted mixing.
Together, these components redefine Krea 2 as an exploratory medium. Instead of returning a single answer, the model is designed to expose a space of possibilities and give users practical ways to move through it using both text and image-based control.
Preserving Knowledge of Real Entities
A subtle but important capability for any image generator is the ability to faithfully represent known entities, people, places, and objects that users might reference simply by name. Krea worried that standard sampling methods could accidentally drop rare or important concepts during data curation.
To prevent this, the team ran PageRank over English Wikipedia, retained the top 90 percent of articles by rank, filtered out concepts that cannot be meaningfully depicted, and then verified coverage across its caption dataset, prioritizing images whose captions referenced rare concepts. The team confirmed afterward that no concepts present in the initial dataset had been dropped entirely from the final training sample.
A Competitive Open Frontier for Image AI
Krea 2's arrival intensifies a crowded open-weights image generation landscape. The company positions its model as "among the top 10" on the Artificial Analysis text-to-image leaderboard and "second place among models from independent labs," a claim that, if it holds under broader community scrutiny, would make K2 one of the strongest openly available image generators.
The technical report, which runs nearly 12,000 words and details everything from data curation principles to distributed training infrastructure, also serves a strategic purpose. By publishing the methodology behind a competitive model, Krea is inviting scrutiny, reproduction, and improvement, the currency of credibility in the open research community.
For creators and developers, the upshot is a capable, openly licensed image model that prioritizes creative range over a single safe default. With the weights available on Hugging Face and the code on GitHub, Krea 2 lowers the barrier for anyone who wants to build, fine-tune, or simply experiment with a state-of-the-art image generator on their own terms.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
