Krea 2
A solid pick for self-hosting and full data control.
Overview
Krea 2 is Krea AI's first in-house foundation model, announced May 12, 2026, then released as open weights in late June. It's a 12.9-billion-parameter Diffusion Transformer trained from scratch, with a deliberate stance: aesthetic diversity over the "default AI look" most generators converge toward. Notably, the training corpus completely excludes AI-generated images, filtered via in-house classifiers — the team found that even a small share of synthetic data caps model quality. The open-source release follows a "train on Raw, generate with Turbo" pattern: Krea 2 Raw is the undistilled checkpoint, ideal as a LoRA training base to capture a style or product; Krea 2 Turbo is the inference checkpoint, generating 2K images in about 2 seconds on consumer hardware. For a brand, that means training an adapter on your own assets and serving it locally, with no dependency on a closed API. Style control works through moodboards and reference images (palette, texture, lighting, composition) rather than endless prompts. Access via the Krea platform (free tier, then $9/month Basic, $35 Pro, $70 Max), the Krea API, or self-hosted via the Hugging Face weights under a community license (commercial use allowed for individuals and small teams, separate enterprise licensing).
Skill profile
Not disclosed
Strengths
- Has a free tier
- Open-source and self-hostable
Limitations
- No native GDPR guarantee
Who is it for
- you want to control cost or self-host
- you handle sensitive EU data
Ideal use cases
- Art direction and style exploration via moodboards
- Consistent brand visuals (LoRA fine-tuning on your assets)
- High-throughput generation (2K in ~2 seconds with Turbo)
Access & availability
Key specifications
Privacy
Advanced data · for expertsArchitecture, modalities, detailed cost, full benchmarks▾
| Arena Elo | — |
| MMLU | — |
| GPQA | — |
| HumanEval | — |
| SWE-Bench | — |
| MATH | — |