← Back to feed

Krea AI launched KREA-2 in early June 2026 as its first foundation image model built entirely from scratch. The medium variant debuted at No. 6 on the Artificial Analysis Text-to-Image Leaderboard. Both medium and large variants entered the Arena.ai Text-to-Image Leaderboard on June 5, with the large model posting 1,121 ELO and the medium at 1,101.

Unlike every other major image generation launch of 2026, KREA-2 launched without claiming to top any benchmark. The positioning is deliberate.

Numbers

• Artificial Analysis Text-to-Image Rank: No. 6 (medium at launch)
• Arena.ai ELO: 1,121 (large), 1,101 (medium)
• Pricing: 10 credits per image (medium), 15 per image (large); add 5 credits when using a style reference

The Design Choice

Most 2026 image models compete on the same metrics: Arena ELO, prompt accuracy, text rendering. KREA-2 does not. The design objective is creative direction: how well the model sustains an aesthetic intent across multiple reference images, not how literally it interprets a single prompt.

In independent testing across portrait photography, fashion editorial, luxury advertising, concept art, and moodboard consistency tasks, KREA-2 scored 9.1-10 out of 10 in creative direction categories. The specific results:

• Fashion editorial: 10/10 for moodboard consistency
• Portrait photography: 9.6/10 for lighting and atmosphere
• Overall creative direction: 9.3/10

Where it underperforms: photorealistic accuracy, text rendering, product consistency, and technical illustration. Midjourney beats it on raw prompt-following. GPT Image 2 beats it on text-heavy projects. KREA-2 beats both when a designer needs to maintain visual direction across a campaign or editorial shoot.

The model also scored highest in prompt accuracy tests from Techscribe’s DeepEval benchmark across stylistic alignment at 90/100, though inconsistencies in spatial positioning and object count held down its overall score relative to top-ranked tools.

Why It Matters

The 2026 image generation market is fragmenting by task rather than converging:

• GPT Image 2: text rendering, structured graphics, general quality (Arena leader)
• Reve 2.0: raw quality, Arena No. 2
• Ideogram 4.0: typography, open weights, first for logos and text-heavy design
• KREA-2: style control, moodboard consistency, creative direction

That split is the more significant story. Arena ELO measures human preference in blind pairings, which tends to favour photorealism and literal prompt compliance. A model optimised for a creative direction workflow will not top a benchmark built around those signals, even if it genuinely serves a distinct and valuable use case.

KREA-2’s Arena ELO of 1,121 places it outside the top tier. Its AA rank of No. 6 and independent review scores suggest the two leaderboards are measuring different things. For designers and creative directors, AA’s more structured evaluation may be the more relevant signal.

What Changed at Krea

KREA-2 is Krea’s first model built on its own weights. All previous Krea products used third-party foundation models with Krea’s tooling layered on top. Building from scratch means the company now controls its own model roadmap rather than depending on external providers for the underlying generation capability. It is a structural shift with long-term implications for product differentiation, pricing, and speed of iteration.