Krea 2 Open Weights: Two Checkpoints, One Practical Workflow

Try GPT Image 2 editor free online in PicEditor.

Krea has released Krea 2 as an open-weight image-model family, giving developers and artists two downloadable checkpoints instead of a single all-purpose file. The release matters because it pairs a flexible foundation model for training with a distilled model for fast generation—without forcing users to choose between customization and practical inference speed.

Those checkpoints are Krea 2 Raw and Krea 2 Turbo. They are related and designed to share LoRAs, but they have not been merged into one model. Understanding that distinction is the key to understanding the release.

What Krea 2 Is

Krea introduced Krea 2 as its first foundation image model built from scratch, with an emphasis on aesthetic range and creative control. Instead of optimizing only for a clean, polished default, Krea says it trained the family to cover photography, illustration, painting, experimental imagery, film grain, motion blur, and less conventional visual styles. (Krea 2 introduction)

The open release uses a 12-billion-parameter dense Diffusion Transformer, or DiT, as its image-generating backbone. It combines that backbone with a Qwen3-VL text encoder using multi-layer feature aggregation and the Qwen Image VAE, which compresses images into a latent representation the diffusion model can process. (Krea 2 open-weight release)

Krea’s technical report describes a pipeline spanning pretraining, midtraining, supervised fine-tuning, preference optimization, reinforcement learning, and optional timestep distillation. The company says it excluded AI-generated images from pretraining, while later stages used synthetic preference pairs and specialized reward models. (Krea 2 technical report)

Krea 2 Raw vs Turbo

Krea 2 Raw is the undistilled base checkpoint before the additional post-training and distillation used for Turbo. It retains a wider, more malleable distribution, which makes it the appropriate starting point for domain adaptation and LoRA training. Raw can generate images directly, but its recommended 52-step sampler makes it considerably slower.

Krea 2 Turbo is the production-oriented checkpoint. It has received additional fine-tuning plus guidance and timestep distillation, reducing generation to eight steps with classifier-free guidance disabled. That makes it the sensible default for local image generation when speed matters. (Official Krea 2 repository)

The official advice is memorable: train on Raw and run on Turbo. A LoRA trained against Raw is designed to express well when attached to Turbo. This gives creators a flexible training target and a fast deployment target without requiring two separate LoRA-training runs.

Are the Two Krea 2 Models Merged Together?

No. Raw and Turbo remain separate checkpoints with separate Safetensors downloads on Hugging Face. Some ComfyUI packages list both models in one repository or workflow collection, but that is packaging—not weight merging.

The confusion is understandable because Krea’s technical report does discuss model merging. During training, the team merged domain-specific supervised-fine-tuning checkpoints to produce a generalist SFT checkpoint. It also used checkpoint averaging during pretraining. Those are internal training techniques; they do not mean the released Raw and Turbo weights were fused into one file.

There is another naming trap. Krea’s hosted API launched with Krea 2 Medium and Krea 2 Large. Medium is positioned as faster and more stable, while Large is more powerful and geared toward expressive photorealism. These hosted API variants are not alternate names for the open-weight Raw and Turbo checkpoints. (Krea 2 API announcement)

Krea 2 Benchmarks and What They Actually Show

Krea’s current official repository describes Krea 2 as the highest-ranked text-to-image model from an independent lab on the Artificial Analysis leaderboard. The technical report, published earlier, described it as top ten overall and second among independent labs. The changing rank likely reflects leaderboard timing, so it should not be treated as a permanent score.

Krea has not published a conventional table of open-checkpoint scores across prompt adherence, typography, human preference, and inference cost. Its public ranking and technical ablations do not support claiming that Raw or Turbo universally beats every competing model.

The practical benchmark is clearer: Turbo reduces the recommended sampling path from Raw’s 52 steps to eight. That is a 6.5× reduction in denoising steps, although real wall-clock speed still depends on GPU, precision, resolution, attention implementation, and model-loading overhead.

Krea 2 Pricing and License Terms

The weights are downloadable after accepting the Krea 2 Community License. This is an open-weight release under a custom license, not an unconditional OSI-style open-source license.

As of August 24, 2026, commercial use under the community license is royalty-free only when the user’s total company-wide annual revenue is below $1 million, measured over the trailing 12 months. Organizations at or above that threshold must obtain an enterprise license before commercial use. Distribution also carries naming, notice, attribution, acceptable-use, and content-filtering obligations. (Krea 2 Community License)

Who Should Use Raw—and Who Should Use Turbo?

Choose Krea 2 Raw if you are training a visual style, character, object, or business-specific domain; studying post-training; or building a derivative checkpoint. Its flexibility is the feature, even though inference is slower.

Choose Krea 2 Turbo if you mainly want to generate images locally, deploy a trained LoRA, or iterate quickly at 1K–2K resolution. Its eight-step sampling makes it the practical everyday checkpoint.

Both are large models. The official Hugging Face repositories identify them as roughly 13B-parameter packages, while Krea specifies a 12B DiT backbone, so local deployment still demands substantial accelerator memory unless quantized or offloaded builds are used. The official release also focuses on text-to-image generation; users should not assume every reference-image, moodboard, prompt-expansion, or editing feature in Krea’s hosted product is contained in the downloadable checkpoint.

Verdict

Krea 2 is an unusually coherent open-weight release because its two checkpoints solve different parts of the same workflow. Raw preserves the adaptability needed for training; Turbo converts that work into fast eight-step generation. Their LoRA compatibility connects the two, but they remain separate models.

For creators who value aesthetic range, custom training, and local inference, Krea 2 deserves testing. The weights are large, the license imposes a commercial revenue threshold, and official benchmark evidence remains limited. Treat it as a customizable image-model family—not one mysteriously merged checkpoint or a license-free replacement for every hosted feature.