Skip to main content

MAI Image 2.5

Microsoft's MAI image model — photorealistic text-to-image plus precise single-image edits, with class-leading in-image text rendering.

Capabilities

FeatureSupport
Text-to-ImageYes
Image-to-Image (editing)Yes
Resolution1K class (up to 1365px long edge)
Reference Images1 (the image being edited)
Aspect Ratios1:1, 4:3, 3:4, 3:2, 2:3, 5:4, 4:5, 16:9, 9:16
SeedNo
Negative PromptNo
Inpainting / MaskNo

Text Rendering

MAI Image 2.5's standout strength is legible, correctly-spelled text inside the image — signage, packaging, UI mockups, posters. Quote the exact copy in your prompt:

A vintage tin coffee sign that reads "MORNING RITUAL — EST. 1962",
weathered enamel, warm afternoon light.

Editing Mode

Connect one upstream image and describe the change. Two behaviors to know:

  • Output follows the input's size. In edit mode the aspect-ratio selector is hidden — the edited image keeps the source dimensions. Pick your ratio at generation time, not edit time.
  • Edits are precise, not regenerative. MAI is good at "change the jacket to red" without redrawing the whole frame.

Prompting Tips

  • Photorealism is the default register. Lens language ("85mm, shallow depth of field"), lighting terms, and material descriptions land well.
  • Quote text exactly. Anything inside quotes is treated as literal copy to render.
  • One edit per pass. Chained single edits ("now make it night") preserve more of the frame than one prompt with five changes.

Limitations

  • Single reference image — no multi-image composition (use Grok Image Pro or Nano Banana 2 for merges)
  • ~1.3K resolution ceiling — no 2K/4K tiers
  • No seed, negative prompt, or masked inpainting
  • One output per run

See Also