Krea 3

• 🆕 Krea's image editing and inpainting model, released November 2026
• 🎯 Box a region, describe the change, only that area redraws
• ⚡ Krea reports most edits return under 8 seconds at native 2K
• 👁️ Composes from multiple reference images; preserves products and characters
• 🔧 Thinking modes trade deliberation for speed via the job-based API
• 🧠 Diffusion transformer with VLM prompt expansion and speculative decoding
• 📏 Outputs at 1K or 2K across common aspect ratios
• 🏢 Built by Krea.ai, a generative creative-tools company

Krea

Krea is a company building generative artificial intelligence products, with a focus on image generation and creative visual tooling. Its work centers on producing high-quality AI imagery through models designed for both fidelity and speed.

Krea 3 in its editing configuration is the inpainting and image-to-image side of Krea's 2026 image model, sharing weights and endpoint with the text-to-image sibling Krea 3. The workflow is instruction-driven: you supply an existing photo, optionally box a region, and describe the change, and the model redraws that area while leaving the rest of the frame intact. It also composes from several reference images, which Krea positions around keeping products, garments and characters faithful across generations.

The clearest shift from the earlier family is intent. Krea v2 Large, Krea v2 Medium and the latency-tuned Krea 2 Turbo were primarily text-to-image systems; Krea describes Krea 3 as a model trained for editing, with text-to-image retained alongside it. On Krea's own launch page, the company states that most edits return in under eight seconds at native 2K, with consistency of products and characters from one image to the next called out as the headline property.

Krea's technical report states that the system is a diffusion transformer served with an in-house post-trained vision-language model for prompt expansion and a speculator that the company reports reaching roughly 700–900 tokens per second, plus activation quantization, context parallelism, timestep caching and custom kernels to cut sampling overhead. The same report describes training data synthesis leaning on image-to-image pairs built from real images, with inpainting and outpainting tasks generated cheaply to teach faithful preservation of unedited regions.

Krea's developer documentation describes an API exposing reference image lists, selectable thinking modes and resolution settings through an asynchronous job-polling interface.