Krea AI's Krea 2 Hits #6 Globally and Promises Open-Source Weights
Krea 2 lands at #6 on the Artificial Analysis text-to-image leaderboard, beating every model except those from OpenAI, Google, and NVIDIA, with weights coming soon.
- Krea 2 debuts at #6 on the Artificial Analysis Text-to-Image Leaderboard, behind only OpenAI, Google, and NVIDIA.
- It is the #1 ranked model from an independent research lab, beating Recraft, Ideogram, and FLUX variants.
- Krea 2 is Krea's first foundation image model trained entirely from scratch, not co-developed with Black Forest Labs.
- Ships in two sizes: Medium for illustration and anime, Large for photorealism and raw textures.
- Open-source weights are confirmed as coming soon, following Krea's prior FLUX.1 Krea open release.
- API live on Fal, ComfyUI, and agent platforms with LoRA training and moodboard guidance.
Krea AI just planted a flag on one of the most watched scoreboards in generative imagery. Their in-house foundation model, Krea 2, debuted on the Artificial Analysis Text to Image Arena at #6 globally and #1 among independent research labs, with the company also confirming that an open-source release is on the way.
The leaderboard uses Elo scores derived from millions of blind preference votes, where users compare two images generated from the same prompt without knowing which model created each image, and higher Elo scores indicate a model is preferred more often by users. Krea 2 Medium debuts at #6, trailing only models from OpenAI, Google, and NVIDIA. The top of the chart is currently held by GPT Image 2 (high) at 1339 Elo, followed by GPT Image 1.5, Nano Banana 2, NVIDIA's Cosmos3-Super, and Nano Banana Pro.
Built from scratch, with a different objective
What makes the placement interesting is that this is Krea's first attempt at training a foundation model end-to-end. Krea 1 was developed in collaboration with Black Forest Labs, while Krea 2 is the company's first image model family trained entirely from scratch.
The model also optimizes for something different than the leaders above it. While GPT Image 2 and Nano Banana chase prompt fidelity and text rendering, Krea 2 is trained to understand how images feel, not just what they contain, so any prompt produces output that is more stylish, more considered, and more visually alive, built for creatives who care deeply about aesthetics. In practice that means the model leans into compositional judgment: color grading that reads as golden hour, intentional depth of field, and framing choices the prompt did not explicitly request.
Two sizes, two personalities
Krea ships the model in two variants that share the same API surface but behave noticeably differently:
- Krea 2 Medium - A smaller, faster, more cost-efficient model with extensive post-training applied, producing stable and consistent outputs. Medium is particularly strong on illustration, anime, painting, and expressive or artistic styles.
- Krea 2 Large - More than twice the size of Medium, with softer post-training applied, giving its outputs a rawer, more textured character. Large is the more powerful model overall and shines on photorealism and raw looks like motion blur, grain, and low dynamic range.
Interestingly, it is Medium, not Large, that landed at #6 on the public arena. Both variants accept up to 10 weighted reference images with positive and negative guidance, plus a creativity slider that controls how literally the model follows the prompt versus injecting its own interpretation.
Speed and control as first-class features
Krea is pitching the model around iteration speed for creative workflows. K2 generates outputs in 15 seconds or less, and thrives on ambiguity: simple prompts yield a wide range of high quality images, not just the same concept repeated four times. That diversity-on-vague-prompts behavior is unusual; most diffusion models collapse onto a single mode when prompts are short.
The model is already wired into the broader ecosystem. The Krea 2 API is available on platforms like Fal and ComfyUI, through agents like Hermes from NousResearch, and with full support for Claude, Codex, or OpenClaw. Custom LoRAs are also supported, letting you fine-tune a style, character, or object from a small image set and stack multiple LoRAs with strength control.
Where it fits, and where it does not
The honest tradeoff is prompt adherence. Krea 2 is more interpretive. It is not bad at following prompts but it applies more aesthetic judgment to them. For creative work where you want the AI to bring something to the image - mood, lighting choices, compositional flair - this is a feature, not a bug. But if you need precise control over every element, Krea 2 will sometimes surprise you in ways you did not want.
That maps cleanly to use cases. The model has a particular affinity for editorial aesthetics. Lookbooks, editorial concepts, fashion mood boards are among its strongest use cases. For e-commerce product photography requiring precise background control, GPT Image 2 or a specialized model is a better fit.
The open-source wildcard
The headline detail from the announcement is the promise that weights are coming. Krea has a track record here, having previously released FLUX.1 Krea openly with full model weights and training details. If Krea 2 ships under similar terms, it would put a model competitive with closed offerings from OpenAI and Google directly into the hands of anyone with a GPU, alongside FLUX.2 and the HiDream family as the strongest open-weight options in the space.