Krea 3

Positioning Overview

krea-3/text-to-image is Krea's model line for text-to-image scenarios, called through the unified HiAPI.ai interface. It accepts text prompts to generate images and offers multiple size presets including square, portrait, and landscape, outputting PNG or JPEG. The line supports two variants, berry and jam: the former leans toward image quality, the latter toward speed. It also allows setting reasoning intensity, negative prompts, and a fixed random seed for reproducible results. It suits development teams that need to balance image quality and generation speed per project and integrate batch generation into their own workflows.

Core Highlights

• Selectable quality and speed tiers——variant offers two options, berry and jam; the former emphasizes quality, the latter emphasizes speed, and you can switch between them based on the task.
• Adjustable reasoning intensity——reasoning_level can be set to low or medium, controlling the degree of reasoning and expansion applied to the prompt before generation; medium takes longer.
• Controllable output size and format——six built-in presets: square, portrait 4:3/16:9, and landscape 4:3/16:9, up to 1024×1024, with output available as PNG or JPEG.
• Reproducible with a fixed seed——passing seed lets identical inputs reproduce the same result, making it easy to debug prompts and compare versions.

Capabilities

Text prompt to image generation: a text description is the only required input and returns a single image; the prompt can be up to 5000 characters, accommodating fairly complete descriptions of subject, style, and composition.

Negative prompt to exclude elements: use negative_prompt (up to 1024 characters) to specify content you do not want to appear in the image, adding another dimension of image control.

Six size presets and format selection: choose square, portrait, or landscape presets according to the final placement, and output PNG or JPEG directly, reducing subsequent conversion steps.

Parameterized result reproduction: once parameters such as seed and reasoning_level are fixed, the same input reliably yields the same output, making it suitable for establishing a repeatable generation baseline.

Use Cases

• Batch image generation for content teams——choose the preset matching the column size and produce one image per call, for placements such as article headers and social media images that need stable output.
• Prompt tuning for designers and developers——fix the seed and adjust prompt and negative prompt item by item, compare output differences, and build reusable prompt templates.
• Producing portrait and landscape assets from the same source——use the same description and switch between portrait and landscape presets to get same-style assets for mobile and web.
• Prototype validation sensitive to generation speed——choose the jam variant and lower the reasoning intensity to quickly get multiple directions early in a project, then use berry to produce refined versions.

Limitations and Selection Advice

Only text-prompt-to-image is supported; it does not accept reference images and does not perform local editing, so for image-to-image or region-specific modifications you should choose a model with the corresponding capabilities. The output limit is 1024×1024, and high-resolution final assets need separate upscaling. negative_prompt only takes effect with the berry variant; when jam is selected, this parameter has no effect. Each call returns one image, so multiple options require multiple requests. If a task requires strictly reproducing an existing character or product appearance, pure text descriptions make consistency difficult to guarantee.