Krea 发布图像编辑模型 Krea 3，2K 编辑 8 秒内完成

Krea has launched Krea 3, an image-editing model that takes a photo or set of reference images and changes them based on written prompts. The company says high-quality 2K edits return in under 8 seconds end to end, letting designers try a new background, outfit, or lighting setup right away instead of waiting for each render.

Krea 3 can edit a selected region of an image and leave the rest of the frame untouched, while bounding boxes let users decide where new elements are placed. Each edit can draw on up to six reference images, including style references. Krea also says it is the first editing model that can be fine-tuned with LoRAs on a specific product, person, or style, so a brand's packshot or a recurring character stays the same from one image to the next. Output is native 2K, with support for up to 4K for print, large screens and campaign assets.

Krea is aiming the model at teams that edit images at volume. In the company's examples, e-commerce teams turn a snapshot into a studio packshot or generate colorways, marketing teams localize a poster into another language or extend one ad into a series, and architects test materials and light on a space before it is built. Pixelcut, a photo-editing app used by small businesses, has been testing Krea 3 with its own users for product photography, and Doodles also uses the model.

Krea 3 is available now on krea.ai, through the Krea API, and via distribution partners including fal, Together AI, OpenRouter, Replicate, Baseten, Hedra, and Nous Research. Pricing is $0.03 per image at 1K and $0.06 per image at 2K, and enterprise teams can request access through Krea's partnerships team.

Krea is a San Francisco AI lab founded by Victor Perez and Diego Rodriguez that trains its own image models and runs a creative workspace combining them with models from other labs. Krea 3 follows Krea 2, the company's first image model built from scratch, which launched in May and was released with open weights in June. Krea says more than 35 million people use its platform, including teams at one in three Fortune 500 companies, and it has raised over $80 million from investors including Bain Capital Ventures, a16z, and Abstract.