Krea 2 is a 12-billion-parameter image model that Krea trained from scratch and released as open weights on June 23, 2026. Krea built it for aesthetic range: its technical report argues that other image models converge on a narrow set of default looks. It reads plain sentences through a Qwen3-VL text encoder and ships in two versions. Raw is the undistilled base, meant for training LoRAs. Turbo is distilled to 8 steps for generating, at 1K to 2K. Krea's advice is to train on Raw and run on Turbo.
What it needs: Krea publishes no VRAM figure, in the model card, the repo or the technical report. ComfyUI recommends its fp8 build for most users: 13.1 GB of weights plus a 5.2 GB fp8 text encoder. That runs on our 12 GB RTX 4070 Ti, in about half a minute per image.
What it's good for: pictures where the look matters, in any medium: film photography, painting, illustration, posters with lettering, product and interior shots. Commercial use is free under the Krea 2 Community License only for companies earning less than $1 million a year. Above that, Krea sells an enterprise license.
In FlixML: `krea2_turbo` (text-to-image, with up to two LoRAs).