Krea 3 edits images in under 8 seconds at native 2K for $0.06

Speed and price as the main pitch

Krea argues that image editing has reached production quality, but that the best models have become slow and expensive. The company says a minute and a quarter of a dollar per edit is acceptable for one hero shot, yet adds up quickly for a full catalog or for testing many ideas in a short time.

Krea 3 is built for the other side of that trade. Most edits come back in under 8 seconds, end to end, at native 2K. Pricing is $0.06 per image at 2K and $0.03 at 1K. The model accepts up to 6 reference images per edit.

Krea compares the model with Nano Banana Pro and GPT Image 2. For the competing models, the company used the median time from request to finished image for edit jobs on Krea between Sep 24 and Oct 1, 2026, queue included. Prices are list prices per image, without image references.

• Krea 3: under 8 seconds, $0.06 per image at 2K
• Nano Banana Pro: 33 seconds, $0.135 per image
• GPT Image 2: 90 seconds, $0.22 per image

In a separate chart, Krea lists Nano Banana 2 at 27 seconds. For 100,000 edits a month at list price, the company puts the cost at $6,000 for Krea 3, $13,500 for Nano Banana Pro and $22,000 for GPT Image 2.

Changing one thing and leaving the rest

Krea says the hardest part of an edit is everything the user did not ask for. Krea 3 is designed to leave untouched areas alone. Change the setting around a product and the label stays sharp, according to the company. A packshot on white can become a campaign photo without the product drifting, which Krea presents as useful for e-commerce and marketing teams.

One example swaps the sofa, chairs and table in a room using three product references, while the rug, the view and a woman reading in the corner remain. Krea says this matters when a single product or label must be updated in an otherwise approved image.

The model also aims to keep people and characters consistent. Given a few references, it can place the same cast in a new street, pose or light. Krea says this separates one good image from a usable set in fashion and film. Style references can be added for a consistent look across a series.

Krea also says the model follows detailed briefs. In one example, it turned a line drawing and a flat palette of six colors into a finished room, with the camera and architecture unchanged. It inherits what Krea calls Krea 2's eye for style, so it works on paintings, vector art and ink drawings while keeping them in their own medium. Sketches, moodboards and old polaroids can also serve as the starting point for a finished photo.

Prompting tips, API and early feedback

Krea recommends starting a prompt with what should stay, for example keeping the product exactly as it is and naming the label or logo. The change is then described the way a photographer would: scene, light and camera. With several images, each should be assigned a role, such as image 1 being the room and image 2 the sofa. The word "only" helps limit an edit to one thing.

Krea 3 is available in the Krea API. Krea lists use cases such as turning a phone snapshot of a product into a studio packshot or a set of colorways, localizing a poster or adapting an ad to a new format, and testing materials and light on a building before it is built.

Pixelcut tested early versions. Its co-founder and CEO Dominique Yahyavi said the product shots "look like they were lit by a professional photographer, not generated by AI." Justine Moore, a partner at a16z, said that when a top-quality edit comes back fast enough to try the next idea, "the model stops being a step in the workflow and becomes the workflow."

Availability and what comes next

Krea 3 is live on krea.ai for the company's 30 million users, and through fal, Hedra, Together AI, Nous Research, OpenRouter and Replicate. Enterprise teams can request access through Krea Partnerships.

Krea admits the model is not real time yet. Its fastest setup finishes an edit in about 4 to 5 seconds, and the company wants to lower that. It also plans to look at architectures that handle high resolution more cheaply, noting that training at 2K was one of the biggest wins of this cycle. A further goal is to let users give instructions by circling an area or writing on the image, and Krea says it is building training data for that now.

For teams that edit images at volume, such as catalogs or ad variants, the pitch is lower cost per image and faster iteration. The speed and price comparisons come from Krea's own measurements and list prices.