Krea 3 – An image model built for editing at scale

The beta launched October 15 and returns most edits in under 8 seconds at native 2K. Krea's comparison table lists Nano Banana Pro at 33 seconds and GPT Image 2 at 90 seconds. Each edit accepts up to six reference images, and a bounding box limits changes to one region. The model is also offered as an API, with 1K output at $0.03 per image. Krea says it is not a distilled copy of a larger model. It was trained for editing and finds the changed region first, then renders only that area. Krea also calls it the first editing model that can be fine-tuned with LoRAs on a specific product, person or style. Distribution already extends beyond Krea's own app: fal lists three endpoints, covering edit, text-to-image and a realtime mode on a 768×768 frame. Pixelcut has been testing it with users, and Doodles also uses it. The speed and quality figures come from Krea. The fal-side timings in the company's demo are pre-generated outputs, not a live benchmark. For founders building on image editing, the competition here is on unit cost and latency, not only quality. At $0.06 per image, high-volume uses such as product photography, ad variants and localized posters become cheaper to run. LoRA fine-tuning could also make brand consistency something a team can own rather than rent.