Scenario MAI Image

Overview

MAI Image 2.5, Microsoft's image family on Scenario, splits into generation members (2.5, 2.5 Pro) and instruction editors (2.5 Edit, 2.5 Pro Edit). The family trait is typography: in-image headlines, labels, and taglines come back legible and placed where the prompt put them: the pick for posters, packaging, covers, and key art that carry real copy. Discover members with search and treat model_schema_get as the contract: the four agree on almost every field and disagree on the one that carries the source image.

Connection and the core loop: see the scenario skill in this repo; model-agnostic image work (sizing families, masks, upscales): the scenario-image skill. If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add scenario-labs/skills --skill <name>); unattended, proceed from tool schemas and flag the gap.

Quick reference

At authoring time (names from the live schema):

Member	Makes	Source input	Ratios
2.5	txt2img	none	11, incl. 21:9
2.5 Pro	txt2img	none	8
2.5 Edit	img2img	referenceImages, array of one	8
2.5 Pro Edit	img2img	image, single file	8

All four share prompt (4096 characters on 2.5 and Edit, 5000 on the Pro pair), aspectRatio (an enum defaulting to auto), and numOutputs (1 to 4, each added image adds cost). No width, height, seed, negative prompt, or mask field exists: aspectRatio is the only sizing control, and edits target elements by naming them in prose. On the generators auto picks a ratio from the prompt; on the editors it matches the source, so leave it there unless re-framing is the point.

Typography is the lever

Write prose sentences, not tag lists: the family reasons over subject, composition, materials, lighting, and mood in that rough order. Wrap exact in-image copy in quotes so it renders verbatim, and give each text block a role, a style, and a place ("the headline 'GAME DAY' in bold condensed type across the top third"). Unquoted or vague copy comes back paraphrased or mangled, and many small text blocks compete, so group them. At authoring time text rendering was tuned for English and outputs landed near 1K, so plan a downstream upscale for print or hero use. Steering is positive-only: explore with numOutputs 2 to 4 and refine the sentence rather than hunting for a seed.

Editing: preserve first, one change

Both editors take one source image and a plain instruction, under different fields: Edit wants referenceImages with exactly one asset id in an array, Pro Edit wants image as a single file. Porting a parameter block between them breaks the run, so re-read the schema when switching. Structure the prompt as what must stay unchanged, then one change, naming the exact element and its exact new state, with replacement copy in quotes ("change the sign to 'OPEN 24/7', same font and color"). Chain passes for several changes; edits hold identity well across iterations. The two shared one public image-edit arena entry (top 3 at authoring time) while a Pro Edit asset cost about four times an Edit asset, so dry_run both and start with Edit.

Worked example: a campaign poster, then a copy swap

1. search with target="models", query="mai image", public=true. Note the generation and edit hits, e.g. model_microsoft-mai-image-2-5-pro and model_microsoft-mai-image-2-5-edit (live hits at authoring time: re-discover each session).
2. model_schema_get on the generation pick: ratio list, prompt cap, defaults.
3. model_run with dry_run=true and parameters={"prompt": "A photorealistic poster of a climber on a granite wall at dawn, warm rim light, the headline 'ASCEND' in tall condensed sans-serif across the top, a small tagline 'Hold your line' lower left, editorial sports aesthetic.", "aspectRatio": "2:3", "numOutputs": 4}. numOutputs moves cost, so re-estimate after changing it.
4. Repeat with wait=false, then jobs_wait with the returned job id, re-called with pending_job_ids on timeout, never a second model_run.
5. asset_display the four outputs and keep one asset id.
6. model_schema_get on the edit pick, then model_run with parameters={"referenceImages": ["asset_x"], "prompt": "Keep the climber, lighting, and layout unchanged. Change only the headline to 'ASCEND HIGHER', same font, color, and placement."}; jobs_wait, then asset_display.

Common mistakes

• Carrying one editor's source field to the other: referenceImages (array of exactly one) and image (single file) are member-specific shapes.
• Leaving in-image copy unquoted: only quoted text renders verbatim.
• Stacking several edits in one instruction: one change per pass, then chain.
• Sending pixel sizes or a ratio the member lacks: aspectRatio is an enum and the lists differ (21:9, 5:4, and 4:5 lived on one member at authoring time).
• Re-running for a pixel-identical result: no seed exists; keep the winning asset and edit it forward.
• Shipping the 1K output to print: upscale downstream (see the scenario-image skill).