Krea 3

Krea 3 is an instruction-following image editor that also generates pictures from text. Start an edit with the change you want, then add one sentence about what should stay the same. A product recolor, a new background or an object moved within a scene is easier to describe as a concrete instruction than as a completely new image brief.

You can combine up to eight input images. Give each a clear role and refer to it by what it shows: the sneaker in the second image, the person in the first. Subject pictures and recipe anchors use the same eight seats. With no references attached, the prompt describes a new picture instead.

The quality selector offers 1K Fast, 1K High, 2K Fast and 2K High. The default is 2K Fast. Fast suits everyday requests; High spends more time planning complex or multi-step edits. Resolution sets the output size, while thinking mode changes the planning effort, so choose each combination for the work it needs to do.

References add a charge based on their pixel area. Kuro caps each at two megapixels before sending it. Use a clear original when faces or small lettering matter, and avoid enlarging a thumbnail. Regular edits can change details beyond the requested area; check the result before using it for another edit. The provider's experimental regional-edit switch is not offered here.

Describe what you want in your own words and Kuro writes the prompt for you. If you would rather write it yourself, this is what Krea 3 responds to.

Write a plain-language instruction. For an edit, lead with what should change, then add one sentence about what should stay the same.

When combining images, identify each by what it shows and by its position: "Put the sneaker from the second image on the person in the first image."

With no images attached, describe the picture you want made.

The quality setting pairs a resolution with a thinking mode. Fast suits everyday requests; High plans complex or multi-step edits more carefully.

What it can't do

• Up to eight input images, shared with saved Subject pictures and recipe anchors.
• Output tops out at 2K.
• A regular edit can change pixels outside the requested area. Say what to preserve, and check the result before chaining another edit.
• Small text and faces need clear input detail. Use the best source you have rather than an enlarged thumbnail.
• The provider expands prompts differently from run to run, so the same instruction does not guarantee the same image.
• Regional edits, negative prompts and seeds are not offered.

Use something else when…

• You need more than eight references: Nano Banana 2.1.
• You need a vector deliverable: Recraft V4.1 SVG.
• The final picture must be larger than 2K: Nano Banana 2.1.