Krea Opens Krea 2 API to Developers Challenging OpenAI on Style Fidelity
Krea opens its first in-house foundation image model to developers, with tunable creativity, style transfer, and moodboard support via REST API
Takeaways
- Krea launched a public API for Krea 2, its first foundation image model trained entirely from scratch, focused on aesthetic diversity and style control.
- Two variants available: Krea 2 Medium ($0.030/image, fast, consistent) and Krea 2 Large ($0.060/image, more powerful, better photorealism).
- Key features include tunable
creativityparameter (raw/low/medium/high), multi-image style transfer with per-reference strength, and moodboard support. - Krea 2 Medium ranks #6 globally on the Artificial Analysis text-to-image leaderboard, #1 among independent research labs.
- Available now via fal.ai, ComfyUI, Runware, and directly through the Krea API; integrates with Claude, Codex, and Hermes agents.
- Main limitations: unreliable text rendering within images and reduced precision for exact product photography specifications.
Krea 2 is now available through a public API. The model, which Krea built entirely from scratch, is the company's first in-house foundation image model and its clearest statement yet that it wants to own the full creative stack, not just wrap other people's models. The API launch puts it directly in the hands of developers who want to build style-aware image generation into their own products.
A different bet on what image generation should be
AI image generators have reached a strange point. Many of them are powerful, fast, and visually impressive, but their outputs can start to feel familiar: the same polished lighting, the same cinematic glow, the same "AI image" texture that looks good at first glance but becomes easy to recognize. Krea's answer to that problem is Krea 2.
Krea 2 is their first foundation image model, trained completely from scratch to give users full control over the look, feel, and creative direction of every image they create. What differentiates it at the training level is data curation and objective function: Krea trained it heavily on aesthetically curated visual content, including editorial photography, fine art, design, and concept work, and tuned the reward signals to prioritize visual harmony, coherent lighting, and style consistency over raw prompt adherence. The result is a model that interprets prompts through an aesthetic lens rather than a literal one.
Krea 2 Medium has already debuted at #6 on the Artificial Analysis Text-to-Image Leaderboard, trailing only models from OpenAI, Google, and NVIDIA. Notably, Krea 1 was developed in collaboration with Black Forest Labs, making this the first model family Krea trained entirely on its own.
What the model actually does well
Most image models handle complex prompts well, but tend to fall short on style, defaulting to something sharp, polished, and "safe." That narrows the creative range you can get out of them. With Krea 2, the team put as much effort into how the model understands style as how it understands prompts. The result is a model that can render a wide range of aesthetics: expressive, raw, niche, experimental, without falling into AI looks.
The model ships with four headline capabilities:
- Aesthetic diversity: Krea 2 can generate an incredibly wide range of styles, with particular strength in artistic and expressive directions that other models tend to flatten out.
- Style transfer: Krea 2 ships with an advanced style transfer system. Developers can pass in a single reference image or combine several, and Krea 2 extracts the style and applies it to the output with precision, with per-reference weighting that lets users decide how strongly each input shapes the final image.
- Creativity parameter: Creativity is a first-class parameter in Krea 2, not an afterthought. At high values, the model takes more creative liberty with style, mood, and aesthetic depth. At low values, it stays close to the literal prompt. At
raw, it renders only what has been explicitly described. This gives developers a meaningful dial between faithful execution and generative interpretation. - Moodboards: A moodboard is the most precise way to set a visual direction with Krea 2. Moodboards are built from a group of images that share an overall creative direction, covering palette, texture, style, mood, or composition. You can create one in Krea and reference it when you run inference through the API.
In a four-model style-transfer benchmark run by Contra Labs, Krea 2 Large ranked second for style fidelity and was only 0.14 points behind GPT Image 2. Krea 2 Large was the only non-GPT model in the evaluation to clear the 3.0 style-fidelity threshold. Krea 2 Large received the top style-fidelity rating on 23% of outputs, while GPT Image 2 received the top rating on 22%.
Two variants, different characters
The Krea 2 API offers two model variants: Krea 2 Medium and Krea 2 Large. Both support the same parameters and feature set, with differences coming down to model size, training stage, and the character of their outputs. If you're not sure which to start with, try Krea 2 Medium: it's faster, cheaper, and handles the broadest range of use cases reliably.
- Krea 2 Medium is a smaller, faster, and more cost-efficient model with extensive post-training applied, making its outputs especially stable and consistent across generations. It is particularly strong on illustration, anime, painting, and other expressive or artistic styles.
- Krea 2 Large is more than twice the size of Medium, with a softer post-training applied, giving its outputs a rawer, more textured, and flexible character. Large is the more powerful model overall and at its best produces results Medium cannot match. It shines on photorealism and "raw" looks like motion blurs, grains, or low dynamic ranges.
Where it falls short
K2's main limitations are moderate performance on complex multi-part instructional prompts, less reliable text rendering within images, and reduced precision for technical or product photography use cases where exact specifications matter. Krea 2 is the weakest of the three major models for text accuracy and is not recommended for use cases where legible text is required. If you need a model that will faithfully execute a 10-clause prompt describing an exact product shot, GPT Image 2 is still the safer bet.
Getting started with the API
You can access Krea 2 through API partners such as Fal, Comfy, or Runware. Krea is also launching with Nous Research, offering access to the Hermes agent from day 0, with full support for Claude, OpenClaw, and Codex in the docs.
The API itself is straightforward. You POST to one of two endpoints, poll a job ID for status, or pass a webhook URL to get notified on completion. Here's a minimal Python-style example of the request shape:
import requests, os, time
API_TOKEN = os.environ["KREA_API_TOKEN"]
BASE = "https://api.krea.ai"
# Submit a generation job
resp = requests.post(
f"{BASE}/generate/image/krea/krea-2/medium",
headers={"Authorization": f"Bearer {API_TOKEN}"},
json={
"prompt": "editorial portrait, soft shadows, film grain",
"aspect_ratio": "4:5",
"resolution": "1K",
"creativity": "medium",
}
)
job_id = resp.json()["job_id"]
# Poll until done
while True:
job = requests.get(f"{BASE}/jobs/{job_id}",
headers={"Authorization": f"Bearer {API_TOKEN}"}).json()
if job["status"] == "completed":
print(job["result"]["urls"])
break
time.sleep(3)
The full parameter set for both variants is identical:
| Parameter | Type | Notes |
|---|---|---|
prompt | string | Required |
aspect_ratio | string | 1:1, 4:3, 3:2, 16:9, 2.35:1, 4:5, 2:3, 9:16 |
creativity | enum | raw, low, medium, high |
image_style_references | array | Optional, with per-reference strength |
moodboards | array | Optional, max 1 currently |
styles | array | LoRA styles (coming soon) |
Pricing
Pricing is per-image and scales with which features you use. Text-to-image starts at $0.030 for Medium and $0.060 for Large. Adding style references adds $0.005, and moodboards add $0.010, but combining both does not stack the surcharge.
- Medium, text-to-image: $0.030
- Medium, style references: $0.035
- Medium, moodboards: $0.040
- Large, text-to-image: $0.060
- Large, style references: $0.065
- Large, moodboards: $0.070
Krea includes compute units on the free tier, so you can make API calls without a credit card. The free plan gives you enough to prototype your integration and test different models. When you're ready for production volume, you can upgrade to a paid plan for more compute and higher concurrency.
Who should actually reach for this
K2 is best suited for creative and editorial work: mood boards, campaign concept exploration, brand identity visuals, fashion and editorial aesthetics, and concept art. It can also help during early campaign exploration. Before committing to a final direction, teams often need to test different moods, color systems, lighting styles, and compositions. A marketing team could explore several visual styles for a product launch: cinematic realism, editorial photography, dreamy illustration, surreal 3D, or retro film. Krea 2's style-reference system makes this kind of exploration easier than relying only on text prompts.
Krea has long been known as a creative AI platform with tools for image generation, video generation, editing, upscaling, and workflow automation. But Krea 2 represents a bigger shift: Krea is no longer only giving users access to external models. It is now building its own model with a clear creative philosophy. An open-source release is also reportedly in the works. For teams building creative tools on top of AI, having a model specifically tuned for aesthetic range, rather than prompt accuracy, opens up a genuinely different design space.