MAI Image 2.6 Photorealistic Image Generation

MAI Image 2.6 is Microsoft's generative image family for photorealistic text-to-image creation and instruction-driven image editing. Atlas Cloud provides flagship and Flash variants through separate generation and editing endpoints.

Use it to produce design-ready commercial imagery, portraits, 3D visuals, and compositions with improved in-image text rendering. For existing assets, the editing models can apply targeted instructions and combine multiple visual references.

Choose the model ID that matches generation or editing and flagship or Flash performance. Submit your prompt and the fields defined in that endpoint's Atlas Cloud parameter schema using your API key.

Select Flash for latency-sensitive or high-volume workloads where lower cost matters. Use the flagship endpoint when you prioritize the family's primary quality-focused model, then choose text-to-image for new visuals or Edit for modifying supplied images.

Text-to-image requests begin with a natural-language prompt, while editing workflows accept an image and editing instructions. Microsoft also documents width, height, model-directed aspect ratio selection, and optional web grounding for this family, but available Atlas Cloud fields should follow the selected endpoint's schema.

Yes. The editing family supports multi-reference composition, and the Atlas Cloud backend confirms that the flagship editor can work across up to five images. Check the exact endpoint schema before sending references because accepted fields can differ between variants.

Atlas Cloud's standard prices are $0.044 per image for Flash text-to-image, $0.047 per image for Flash editing, $0.088 per image for flagship text-to-image, and $0.099 per image for flagship editing. Billing is pay as you go, allowing developers to select an endpoint for each workload.

Microsoft provides the family through hosted services, and the reviewed first-party materials do not publish downloadable weights or an open-weight license. Treat it as a hosted model rather than a self-hosted deployment.

Leaderboard positions reflect particular prompts, voters, evaluation dates, and task categories rather than every production workload. Compare the flagship and Flash endpoints with your own prompts, references, latency requirements, and review criteria before selecting a default.

First confirm the model ID, use the correct generation or editing schema, and validate every required input field. If the request completes but misses your intent, simplify the instructions, quote required text explicitly, identify elements that must remain unchanged, and review the result before production use.