News

Krea 3 Opens in Beta: Six-Reference Image Edits in Under Eight Seconds, and a Fine-Tuning Claim Worth Testing

19 October 2026 · RenderBob team

Krea's 15 October post says most Krea 3 edits return in under 8 seconds at native 2K, from up to six references. The API adds a charge per reference, and LoRAs are marked coming soon.

Krea published Krea 3 on 15 October. The launch post says most edits come back in under 8 seconds, end to end, at native 2K, at $0.03 an image for 1K and $0.06 for 2K, from up to six reference images. It is live on krea.ai and through fal, Hedra, Together AI, Nous Research, OpenRouter, and Replicate. Enterprise access goes through Krea's partnerships team.

What the launch post claims

Krea says an edit can change the scene around a product and leave the label alone, and that several references can hold a person steady across a set. The post quotes Pixelcut co-founder Dominique Yahyavi on the lighting of product shots, and it names Pixelcut as an early tester. The FAQ on that page says the model generates natively at 2K. The post does not mention LoRA fine-tuning, a 4K mode, or a beta.

What TestingCatalog adds

TestingCatalog's article the same day quotes a Krea post that opens a beta and claims quality matching Nano Banana Pro or GPT Image 2.5, plus bounding boxes and multi-turn edits. The article says Krea calls this the first editing model that can be fine-tuned with LoRAs on a product, person, or style, and that output can go to 4K for print. It names Doodles as well as Pixelcut, and it adds Baseten to the partner list. There is no independent test in the piece.

The TestingCatalog text we retrieved has no sponsored label. The beta wording, the quality comparison, the LoRA claim, and the 4K line are that article's account of what Krea said.

What the API says now

Krea's API overview lists up to 8 image URLs. The launch post says six. Resolution is 1K or 2K. Width and height can be requested up to 4096, and the API rounds them onto a 1K or 2K canvas. Set regional_edit to true when the rest of the frame has to stay put. Krea says that mode leaves everything outside the change exactly as it was. The API rejects a loras field with a coming-soon message, and it does the same for style_urls. The page says the fast and high thinking modes do not require beta access.

The API price is the launch figure plus the references. Each 1K reference adds $0.02, each 2K reference adds $0.07, and high thinking adds $0.02. The launch post's comparison table prices the $0.06 figure without image references. With five or more references in one request, each image is resized to at most 1536 pixels on its longest side.

Against the other stills caps

Six on the launch post, or eight on the API, sits under Nano Banana 2.1's 14 and FLUX 3's 10, from our 12 and 15 October comparisons. The figure is the interface limit.

Questions before you use it

• Fine-tuning on a person needs that person's written consent, with scope, duration, training, approval, and payment written down. That is the list in our 14 October note on Nicolas Cage. The API does not offer the LoRA yet.
• Enterprise workspaces can send a zero-data-retention header. In that mode every reference has to be a base64 data URI. For a normal request, ask where the images are stored and who can read them.
• The launch post and the API overview do not say whether outputs carry content credentials.
• Run the same edit on Krea 3, Nano Banana 2.1, and your current tool. Score identity, and score how much of the untouched region stayed.

When that still has to go into a cut, RenderBob lets you choose local or cloud for each clip and keep the references with the studio.

What to watch

A LoRA path that is actually available, with a published price, and a comparison that is not Krea's own timing table.