Krea 3 is an image editing model built for production, according to the source. It changes only what the user asks and leaves the rest of the image untouched. The source says the model finishes an edit quickly and has a per‑image cost, although specific numbers are not provided in the text. Krea 3 is distributed through the world’s leading organizations to help them advance their industries.
A core capability highlighted is regional editing: the user can draw a box around the area to modify and describe the desired change; Krea 3 redraws only inside that box while preserving the surrounding pixels. The source notes that three separate edits of the same image, made with this regional approach, show that changed pixels are measured against the original.
The model can match a specified hex code on a product and apply it only to that product, leaving other colors unchanged. It also can composite elements from different sources—such as placing a person, a knit sweater, and a silver bag into a candle‑lit Art Deco bar scene—by pointing at people, products, places or styles by number and putting them together in one picture.
Krea 3 allows users to alter styling, setting, and mood while keeping a face familiar, enabling exploration of character consistency across a sequence of edits. It renders output at 2K resolution natively, and the source invites users to move a loupe over the result to see the real pixels.
In a side‑by‑side illustration, one image processed with twelve identical prompts was edited by Krea 3, Nano Banana Pro, and GPT Image 2.5. The source reports illustrative playback times of 8 seconds per edit for Krea 3, 26.8 seconds for Nano Banana Pro, and 39.9 seconds for GPT Image 2.5 Sunburst, representing the median request‑to‑result time for each model’s 12‑image batch, including queue time.
Costs are described as approximate USD list prices for 2K edits, excluding promotions, with the note that GPT Image 2.5 pricing remains a draft estimate pending validation. Precision is defined as the share of pixels outside the edit area left unchanged; the source explains that this metric is measured after aligning outputs at quarter resolution.
Krea 3 emphasizes that it is not a distilled copy of a larger model; it spends its compute on the region being changed rather than the whole image, which the source says is appropriate because most edits affect only a small part of the picture. The model is positioned as best for product‑scene changes, object removal, and photo cleanup.
Enterprise customers can run Krea 3 on their own GPUs in the cloud, and the source invites teams to discuss license terms and hardware sizing. The company offers to run a user‑supplied workload alongside their current solution and share outputs, acceptance rate, and cost. Access is available via the Krea API or the Krea app, and volume pricing can be evaluated by contacting the team.