MAI-Image 2.5 Pro Edit API by MICROSOFT

MAI-Image-2.5-Pro Edit (Image-to-Image)

MAI-Image-2.5-Pro Edit is Microsoft AI's highest-fidelity image model applied to image-to-image editing — precise, controllable changes to an existing image driven by a natural language instruction. It uses a diffusion-based generative approach that progressively refines the image, producing strong alignment between the instruction and the result while preserving the original composition and layout. Compared with the standard MAI-Image-2.5, the Pro tier is tuned for robust object consistency, stronger visual reasoning, and deeper world knowledge, so edits hold up in visually dense scenes and across repeated iterations on the same asset.

Released in public preview on July 23, 2026 (model version 2026-06-19), it is the premium tier of the MAI-Image-2.5 family, alongside the balanced base model and the cost-efficient Flash variant.

Key Capabilities

• Surgical object editing — Remove, replace, recolor, or reposition a specific element without disturbing the rest of the image.
• Object consistency across complex scenes — Objects retain the same identity, materials, proportions, markings, and orientation throughout a visually dense composition, so an edit in one region doesn't destabilize another.
• Character consistency across views and moments — A person or character stays recognizably the same across poses, camera angles, expressions, clothing, and lighting conditions.
• Material and physical-property accuracy — Edited materials keep behaving according to their real-world properties, including reflection, translucency, weight, texture, and deformation.
• Spatial and geometric reasoning — Inserted or moved objects sit coherently in three-dimensional space, with credible scale, perspective, occlusion, and structural relationships.
• Inpainting & attribute changes — Fill in missing regions, remove unwanted content, and change attributes of existing subjects seamlessly.
• In-image text updates — Rewrite signage, labels, and packaging text with the accurate text rendering that defines the 2.5 generation.
• Artifact cleanup — Remove motion blur, noise, and compression artifacts.
• Layout adaptation — Reframe, recompose, or restructure a scene while keeping the subject consistent.
• World knowledge — Understands cultural references and complex concepts, so shorthand instructions resolve into plausible edits.
• Product, branding & commercial design — Design-aware refinement for product imagery, marketing visuals, brand assets, and commercial creative workflows.

Request Parameters

Note that the edit endpoint takes no size parameter: output dimensions follow from the reference image rather than being requested explicitly. If you need to control resolution directly, generate with the text-to-image variant instead.

Response

An edit request returns a prediction object with a status of created, processing, completed, timeout, or failed. Poll the returned urls.get endpoint (or use enable_sync_mode) until the task completes; data.outputs then holds the array of edited image URLs, or BASE64 strings when enable_base64_output is set. The response also carries has_nsfw_contents flags per output, executionTime, and a timings.inference breakdown in milliseconds.

Pricing

Pricing for this model has three components: the input text tokens in the prompt, a fixed input image fee, and a fixed per-output-image fee.

Pricing Formula

cost = countTokens(prompt) / 1,000,000 × $5.00 + $0.12 + $0.011

Which simplifies to:

cost = countTokens(prompt) / 1,000,000 × $5.00 + $0.131

For typical prompts of a few hundred tokens the token component is negligible, so the effective cost is approximately $0.131 per edit.

Examples

Best Use Cases

• Hero asset refinement — High-value campaign imagery where an edit must be invisible and the material realism must survive it.
• Product retouching & variants — Swap colors, finishes, or backgrounds from a single base shot while reflections and textures stay physically plausible.
• Marketing asset updates — Refresh text, logos, or seasonal elements in an existing creative instead of reshooting.
• Content localization — Replace in-image text across languages and markets.
• Iterative design — Run many successive edits on the same asset while composition and identity hold across iterations.
• Photo restoration & cleanup — Remove motion blur, noise, and defects from archival or imperfect source images.
• Character-driven series — Extend an existing shot into new poses or moments with the same recognizable subject.

Pro Tips

• State both halves of the edit: what to change and what to preserve (e.g., "change the red car to matte navy blue, keep the background and lighting unchanged").
• For text edits, quote the exact string to replace and the exact replacement.
• Describe the desired end state rather than only the removal — "place the product on a white studio backdrop" beats "remove the background".
• Name materials explicitly when changing a surface, so the model applies the right reflection, translucency, and texture behavior.
• Make one substantive change per request and chain requests for compound edits; this keeps consistency tight and makes results easier to review.
• Supply the highest-quality source image available — JPEG or PNG — since the edit inherits its resolution and detail.

Technical Specifications

Responsible Use

Despite data filtering and system-level content classifiers, image editing models can still produce unexpected or harmful output. Only edit images you have the rights to use, apply content safety mitigations appropriate to your use case, avoid edits that could mislead about real people or events, and disclose that images are AI-generated when publishing them.

Related Models

• MAI-Image-2.5-Pro Text-to-Image — The same Pro model generating images from text alone, with explicit resolution control.
• MAI-Image-2.5 Edit — The balanced base tier for image editing.
• MAI-Image-2.5-Flash Edit — Faster, lower-cost image editing variant.