Krea Launches Krea 3 to Edit Images in 8 Seconds for $0.06

Krea has opened the public beta for Krea 3, a foundation image-editing model built for fast, targeted changes to existing images. The company says 2K edits return in under eight seconds end to end, with output quality intended to compete with Nano Banana Pro and GPT Image 2.5.

Krea 3 accepts a source image, written instructions, and optional reference images for products, people, places, or styles. Its combination of regional controls, low per-image pricing, and short latency targets interactive design tools, e-commerce systems, advertising pipelines, and automated creative workflows.

Krea’s speed and price claims

Krea’s published comparison uses 12 matched prompts across the three models. The figures are vendor-reported, so production teams will need to reproduce them with representative images, concurrency, and network conditions.

At those published rates, a Krea 3 edit costs about 56% less than Nano Banana Pro and 71% less than GPT Image 2.5. Krea charges $0.03 for a 1K image and $0.06 for a 2K image.

Regional edits keep the frame stable

Krea 3’s main technical addition is region-aware editing. Users can select an area to change, place new elements with bounding boxes, and instruct the model to preserve the rest of the frame. A jacket can be changed to cherry-red suede, for example, without intentionally regenerating the person, background, or lighting.

The model also supports multiple visual references. Krea’s launch material describes edits using as many as six references, while the API documentation accepts as many as eight image URLs. Developers should verify the limit for the specific endpoint and mode they use.

Custom LoRA fine-tuning can adapt the model to a product, person, character, or visual style. A LoRA is a small trained adapter layered onto the base model, allowing teams to improve consistency without retraining the entire system. The feature could help catalogs and campaigns place the same product or character into many scenes with less visual drift.

The API surface stays compact

Krea 3 is available through Krea’s site, the early-access API, and providers including fal, Together AI, OpenRouter, Replicate, Baseten, Hedra, and Nous Research.

A minimal request through fal looks like this:

import fal_client

result = fal_client.subscribe(
    "fal-ai/krea-3",
    arguments={
        "prompt": "Change only her coat to cobalt blue",
        "image_urls": ["https://example.com/portrait.jpg"],
        "regional_edit": True,
        "resolution": "2k",
        "thinking_mode": "fast",
    },
)

Because the API remains in early access, production evaluations should verify rate limits, queue behavior, timeout handling, regional-selection controls, and output consistency under concurrent load. Provider-specific schemas and availability may also differ.

Precision is the central quality test

Krea published a benchmark measuring how many pixels outside an edit region remain unchanged. The company aligns each output with its input and scores both at quarter resolution, preventing a size difference from counting as an edit. That method focuses on broad unwanted changes, although reduced-resolution scoring can overlook smaller pixel-level differences.

For the prompt “change only her coat to cobalt blue,” Krea reports that its Precise mode preserves more of the surrounding image than Nano Banana Pro or GPT Image 2.5. Both competing models regenerated portions of untargeted regions in Krea’s examples.

Krea also reports a score of 4.81 out of 5 for its Ultra tier on CPI-Bench, an Alibaba Group benchmark covering real-world image-editing tasks. That result places the tier behind GPT Image 2.5 and ahead of GPT Image 2, Seedream 5 Pro, and Nano Banana on the cited leaderboard. Krea’s materials do not clearly map the Ultra tier to the API’s fast and high modes, so comparisons should record the exact endpoint and settings.

Other advertised capabilities include:

• Native 2K rendering, with 4K output advertised for print workflows
• Hex-code color controls limited to the selected product or region
• Numbered references for composing people, products, places, and styles
• Character consistency across multi-turn editing sessions
• Typography-aware edits for posters, packaging, and covers

Beta limits remain visible

Krea describes text-to-image generation in Krea 3 as experimental. Krea 2 remains the company’s generation-focused model, making it the more established option for workflows that begin without a source image.

The company says the full Krea 3 release and additional benchmarks will follow the beta announcement by one week. LoRA training and style references, also called srefs, require early-access approval through a direct request, and their broader API availability has not been specified.

Eight-second edits change the loop

Latency determines whether image generation behaves like a queued render or an interactive tool. A six-cent 2K result delivered in under eight seconds allows an application to offer rapid revisions, generate several candidates in parallel, or incorporate image calls into automated workflows without adding 30 to 40 seconds at each step.

Krea demonstrates voice-note prompting in which a spoken instruction triggers an edit before the next instruction is complete. Similar latency could support live creative assistants, product configurators, campaign-variant generators, and agents that inspect an output before deciding on the next edit.

Distribution through several inference providers lowers integration friction for teams already using those platforms. Krea has also discussed broader LoRA access and the possibility of releasing model weights, although neither option has a confirmed release schedule.

Benchmark before changing pipelines

A useful evaluation should use production images and measure the behavior that affects the final application:

• Median and 95th-percentile latency under expected concurrency
• Pixel changes outside selected regions
• Product, character, and typography consistency across repeated edits
• Prompt adherence with multiple references
• Retry rates, moderation failures, and provider queue times
• Total cost per accepted output, including discarded generations

Krea 3’s public beta presents a credible option for commercial editing, product photography, marketing creative, and character-driven content. Its practical advantage will depend on whether the reported speed, precision, and pricing hold under each team’s workloads.