Krea 2 Open‑Source Release – Two Variants for Text‑to‑Image Generation

Krea has launched Krea 2, a new open‑source text‑to‑image model that claims to be the most aesthetic model from an independent lab. The release ships in two variants: Krea 2 Raw, a CFG‑guided model optimized for control, fidelity and training, and Krea 2 Turbo, a distilled, few‑step version that can render up to 2K resolution quickly. The models are tuned for natural‑language prompts; longer, specific prompts yield the best results, while short prompts still work well. Text can be rendered in images by wrapping words in quotes, and the model supports a growing library of LoRAs that can be loaded via their Hugging Face path. Krea 2 also includes a set of style LoRAs and a dedicated “Krea 2 LoRA” that can be used to extend or customize the model. The team has published a technical report, code repository, and demo spaces on Krea.ai and Hugging Face, and they are hosting an AMA and a livestream with the ComfyUI team. The release is available immediately and can be integrated into existing pipelines for high‑quality, fast image generation.

The announcement emphasizes the model’s versatility for both creative and training purposes, and the team invites feedback on rough edges. They also highlight the ability to prompt the model as if describing an image to a person, which can simplify prompt engineering. The release includes a new “Krea 2 Raw” variant for fidelity and a “Krea 2 Turbo” variant for speed.

Overall, Krea 2 offers a robust, open‑source alternative to proprietary models, with two distinct use cases and extensive LoRA support.