Krea 2 Pursues Art: The Highest-Completion AI Image Model Excelling at Anime Illustrations
- Krea pursues art with AI
- Krea 2 is the highest-completion image generation AI model to date
- Custom license but generous for commercial use
Introduction
Hello, this is Easygoing.
Today, I'd like to introduce Krea 2, a new image generation AI that excels at anime illustrations.
Krea.ai Pursues Art
Krea.ai is an AI startup based in San Francisco, USA.
The company's CEO, Víctor Pérez, and co-founder Diego Rodriguez are both from Spain. They studied Acoustic and Audiovisual Systems Engineering at a university in Barcelona before transitioning into AI researchers.
The company aims to ensure that "AI does not replace (human) creative workflows, but instead gives humans a faster and smarter canvas," with the biggest theme being how to incorporate "aesthetics/taste" into AI.
Krea 2 Is Their First In-House Developed Model
On July 31, 2025, Krea.ai released their first image generation AI model, Krea 1 (Flux.1-krea-dev), as open weights.
gantt
title Krea Roadmap
dateFormat YYYY-MM-DD
tickInterval 12month
axisFormat %Y
section 1st Generation<br>(SD1_vae)
Stable Diffusion 1 : done, 2022-08-22, 2026-06-30
section 2nd Generation<br>(SDXL_0.9_vae)
Stable Diffusion XL : done, 2023-07-26, 2026-06-30
section 3rd Generation<br>(Flux.1_vae)<br>(wan_vae)<br>(Qwen-Image_vae)
Flux.1 : 2024-08-01, 2026-06-30
Wan : 2025-02-25, 2026-06-30
Qwen-Image : 2025-08-04, 2026-06-30
Z-Image : 2025-11-25, 2026-06-30
Cosmos-Predict2 : 2025-04-30, 2026-06-30
Anima : 2026-01-26, 2026-06-30
Krea 1 (Flux.1-krea-dev) : crit, 2025-07-31, 2026-06-30
Krea 2 Medium : crit, 2026-06-22, 2026-06-30
section 4th Generation<br>(Flux.2_vae)
Flux.2 : 2025-11-25, 2026-06-30
ERNIE-Image : 2026-04-07, 2026-06-30
section 5th Generation<br>(No VAE)
HiDream-O1-Image : 2026-05-08, 2026-06-30
Krea 1 was a fine-tuned version of the Flux.1 [dev] model, which was highly regarded at the time for its expressive power. It was praised for "natural expressions that don't feel AI-like."
Then, on June 22, 2026, Krea.ai proudly unveiled their in-house developed model, Krea 2.
Krea 2 Is the Ultimate Evolution of Image Generation AI!
Now, let's take a look at the Krea 2 model lineup and its features from the Krea 2 technical report.
Krea 2 Model Lineup
- Krea 2 Large: Flagship model, available only via cloud (API)
-
Krea 2 Medium: Standard model, open weights
- Krea 2 Raw: Base model for training
- Krea 2 Turbo: Distilled model for generation
The Krea 2 models come in two grades: the Krea 2 Large (flagship model available only in the cloud) and the Krea 2 Medium (released as open weights).
Furthermore, Krea 2 Medium has two variants: the Krea 2 Raw base model for training and the Krea 2 Turbo distilled model optimized for stable output.
When performing image generation locally, I recommend using the Krea 2 Turbo model in most cases.
Krea 2 Medium Architecture
| Krea 2 Medium | Developer | Model | License |
|---|---|---|---|
| Text Encoder | Alibaba | Qwen3-VL（4B） | Apache-2.0 |
| Transformer | Krea.ai | Krea-2_transformer | KREA 2 Community License |
| VAE | Alibaba | Qwen-Image_vae | Apache-2.0 |
The Krea 2 Medium model, released as open weights, uses Alibaba's Qwen3-VL (4B) as the text encoder and Alibaba's Qwen-Image_vae as the VAE.
The Text Encoder Is the Standard Qwen
During development, Krea 2 conducted detailed comparisons of various components.
Text Encoders and VAEs compared during Krea 2 development (Modified from the Krea 2 Technical Report)
| Component | Baseline | Ablations | Final component |
|---|---|---|---|
| Text encoder | T5-XXL encoder | T5Gemma, Qwen 2.5 VL, Qwen 3 VL, umT5 | Qwen 3 VL |
| Autoencoder | FLUX AE | Qwen Image VAE, DC-AE, FLUX 2 AE, Internal VAE | Qwen Image VAE & FLUX 2 AE |
For the text encoder, Krea 2 initially evaluated T5-XXL, but ultimately adopted Qwen3-VL due to its advantages in visual understanding (VLM) and strong multilingual capabilities.
The VAE Is the High-Quality Qwen-Image_vae
Krea 2 Medium uses Qwen-Image_vae for its VAE. In terms of VAE generations, it belongs to the third generation of image generation AIs, same as Flux.1 and Krea-1.
gantt
title Krea Roadmap
dateFormat YYYY-MM-DD
tickInterval 12month
axisFormat %Y
section 1st Generation<br>(SD1_vae)
Stable Diffusion 1 : done, 2022-08-22, 2026-06-30
section 2nd Generation<br>(SDXL_0.9_vae)
Stable Diffusion XL : done, 2023-07-26, 2026-06-30
section 3rd Generation<br>(Flux.1_vae)<br>(wan_vae)<br>(Qwen-Image_vae)
Flux.1 : 2024-08-01, 2026-06-30
Wan : 2025-02-25, 2026-06-30
Qwen-Image : 2025-08-04, 2026-06-30
Z-Image : 2025-11-25, 2026-06-30
Cosmos-Predict2 : 2025-04-30, 2026-06-30
Anima : 2026-01-26, 2026-06-30
Krea 1 (Flux.1-krea-dev) : crit, 2025-07-31, 2026-06-30
Krea 2 Medium : crit, 2026-06-22, 2026-06-30
section 4th Generation<br>(Flux.2_vae)
Flux.2 : 2025-11-25, 2026-06-30
ERNIE-Image : 2026-04-07, 2026-06-30
section 5th Generation<br>(No VAE)
HiDream-O1-Image : 2026-05-08, 2026-06-30The third generation of image generation AIs features models equipped with Flux.1_vae and its derivatives wan_vae and Qwen-Image_vae. They are characterized by a good balance of image quality and compression efficiency, representing technical maturity.
About VAE Image Quality and Compression Efficiency
Krea 2 Medium initially evaluated Flux.1_vae, but ultimately adopted Qwen-Image_vae, which converges faster than Flux.1_vae without sacrificing reconstruction quality (i.e., better performance as a VAE).
Krea 2 can be described as a model that aims for ultimate perfection by meticulously comparing available components and selecting the best ones.
Krea 2 Turbo Examples!
Here are some example images generated with the Krea 2 Turbo model.
The image generation prompts were created using the official workflow's prompt enhancer.
Flower Field
A dynamic Dutch angle close-up portrait of a white anime-style woman with blonde hair and striking blue eyes, captured mid-turn around in a stylish, elegant light blue dress that glows under a dramatic face spotlight against a stark black background; her facial features are sharply detailed, skin rendered smooth and luminous, with vivid color saturation emphasizing alizarin, burgundy, rouge, crimson, indigo, royal blue, deep blue, and deep purple tones in her attire and surroundings; soft bokeh blur envelops the scene while a single vibrant cluster of pink hydrangeas with green leaves blooms in the foreground, contrasting with a minimalist white house featuring a white door behind her — the garden is neatly tended with scattered pink and white blossoms, all bathed in cinematic lighting that enhances the intense, professional, and eye-catching aesthetic of the composition.Figure Garden
A dynamic Dutch angle close-up of a white woman’s upper body and face, captured in a realistic style with dramatic lighting: her blue eyes gleam under a focused spotlight against a pure black background, emphasizing her features and the vivid, colorful tones of alizarin, burgundy, rouge, crimson, indigo, royal blue, deep blue, and deep purple in her hair, makeup, and elegant sleeveless black top; she turns around with one hand gently resting on her shoulder, long wavy brown hair flowing slightly in the wind, wearing a delicate ring, surrounded by soft bokeh blur and minimalist composition that isolates her as the central subject — while simultaneously, in a separate scene, a serene Japanese alleyway at night glows with warm golden street lamps, traditional wooden architecture lining the path, and a solitary figure walking away from the viewer, bathed in ambient light, both scenes rendered with high realism and atmospheric depth.Street Corner
A dynamic Dutch angle close-up shot of a white anime-style woman with cybernetic hair, wearing an elegant black dress accented with gold detailing, standing against a stark black background under a dramatic face spotlight; her blue eyes gleam with confidence as she turns slightly toward the camera, wind gently tousling her hair, surrounded by vivid bokeh blur and a simple backdrop featuring a cluster of purple hydrangeas with green foliage in soft focus behind her, rendered in rich, saturated colors including alizarin crimson, burgundy, rouge, indigo, royal blue, deep blue, and deep purple tones, with transparent highlights enhancing her stylish silhouette, capturing the intensity of her gaze and the elegance of her pose in a cinematic, high-contrast anime aesthetic.Features of the Krea 2 Model
The features of the Krea 2 model that I noticed are as follows:
- Minimal residual noise (clean black expressions)
- Natural colors
- Accurate sense of three-dimensionality (precise reproduction of camera aperture and depth of field)
The output from the Krea 2 model has very little residual noise, resulting in particularly clean and beautiful black expressions and an overall clean impression.
Image generation AI models using VAEs tend to have imbalanced colors because the AI doesn't truly understand color itself, but Krea 2 achieves natural color expression through thorough tuning.
Additionally, Krea 2 accurately reproduces elements like camera aperture and depth of field, making it feel less unnatural than typical AI outputs and resulting in a natural finish.
Krea 2 is a highly polished AI model that reflects the philosophy of its founders, who come from a background in video engineering.
Custom License but Generous for Commercial Use
Finally, let's check the license for Krea 2.
Krea 2 uses Krea.ai's own KREA 2 Community License v.1.
Key Features of KREA 2 Community License v.1
- Krea.ai does not claim rights to the outputs
- Individuals or organizations with annual revenue under $1 million can use it commercially for free
- Derivative models must implement appropriate filtering to prevent inappropriate outputs and are required to include "Krea" in the model name
The Krea 2 model explicitly permits commercial use for individuals and small businesses.
On the other hand, when releasing derivative models, they must implement appropriate filtering to prevent inappropriate outputs and are obligated to include "Krea" in the model name.
The KREA 2 Community License v.1 seems similar to the Stability AI Community License used for Stable Diffusion 3 when compared to previous models.
I think the Krea 2 model has a generous license that allows commercial use for personal purposes, but since it is a custom license, please be sure to review the full license text yourself before using it.
Summary: Krea Pursues Art with AI
- Krea pursues art with AI
- Krea 2 is the highest-completion image generation AI model to date
- Custom license but generous for commercial use
This time, I introduced the Krea 2 model.
Krea 2 is a very interesting AI model where you can strongly feel the developers' philosophy. Thanks to its solid development approach, I believe it is the highest-completion image generation AI model to date.
The Krea 2 technical report mentions that "the next training will simplify the architecture and integrate various components into a single model," suggesting the development of a next-generation model using the same UiT architecture as the previously introduced HiDream-O1-Image.
About the HiDream-O1-Image Model
Developing a next-generation model with the UiT architecture is still a highly challenging area, but I believe the models created by Krea.ai will undoubtedly raise the industry standard in terms of completion, so I'm looking forward to its release.
Thank you for reading until the end!