MAI Image 2.5: Photorealistic Generation and Editing

Covers MAI Image 2.5 and MAI Image 2.5 Edit

Microsoft's MAI Image 2.5 family on Scenario pairs a text-to-image generator with a natural-language editor. Generate campaign-ready frames from a detailed prompt, then localize palettes, swap backgrounds, restyle layouts, or shift art direction on the same stack. Both models ship through Fal and target marketing, food, fashion, film, and game pipelines where fidelity, embedded type, and controllable edits matter.

The short version

• Generate with MAI Image 2.5: long prompt + aspect ratio.

• Edit with MAI Image 2.5 Edit: one reference image + instruction prompt.

• Put literal text in quotation marks when words must render inside the frame.

• Write prompts with lighting, camera, layout, and mood detail. Short generic prompts underperform.

Which Model Should I Use?

Start on MAI Image 2.5 when you need a net-new frame. Open MAI Image 2.5 Edit when the composition is close but palette, background, headline, or art direction needs a surgical change. On Scenario today, plan on one reference image per edit run; multi-image uploads may fail at the provider even though the schema lists up to twelve slots.

Parameters

MAI Image 2.5 (text-to-image)

• Prompt (required). Up to 4,096 characters. Describe subject, style, mood, composition, lighting, and any text that must appear. Quote exact wording for mastheads, headlines, titles, and labels. Long, specific prompts outperform one-line requests.

• Aspect Ratio. Default Auto lets the model infer proportions from the prompt. Or lock a preset: 16:9, 3:2, 4:3, 1:1, 3:4, 2:3, 9:16, and others listed in the UI. Match the publish target (9:16 for Stories, 16:9 for banners, 2:3 for editorial covers). Prefer 16:9 or 3:4 over 21:9 or 4:5 if a run fails (see Limitations).

• Image Count. 1 to 4 outputs per run.

MAI Image 2.5 Edit (image-to-image)

• Images (required). Upload one reference image per run on Scenario: a photo, render, poster, or illustration. Start from high-quality sources such as MAI Image 2.5 or GPT Image 2 when fidelity matters.

• Instructions (required). Up to 4,096 characters. Describe the edit in plain language: what to change, what to preserve, and any new text to add. Be explicit about elements that must stay untouched (pose, layout, character design, logo placement).

• Aspect Ratio. Default Auto to match the source or infer from the prompt. Presets: 16:9, 3:2, 4:3, 1:1, 3:4, 2:3, 9:16. Edit omits ultrawide 21:9 and 5:4 from the generator list.

• Image Count. 1 to 4 edited variants per run.

How MAI Image 2.5 Works

MAI Image 2.5 is a diffusion-based model tuned for photorealistic output and legible embedded text. On Scenario it is text-to-image only: no reference upload on the generator page.

Microsoft positions the family among top Arena text-to-image and image-editing models at launch (June 2026). Outputs respect a roughly one-megapixel total pixel budget (for example 1024×1024, or wider or taller layouts within that cap).

Verified generation examples

Fashion editorial cover (2:3)

1960s high-fashion magazine cover blending photoreal portrait with fashion-illustration linework. Model in sculptural ivory coat with sharp geometric shoulder wings and sunray pleats across the torso, aloof editorial gaze. Masthead text "MAISON" in bold classic red serif across the top. Date line "SEPTEMBER 1968" upper left. Right column headlines "THE NEW SILHOUETTE" and "ARCHITECTURAL CHIC" in elegant italic serif. Warm aged paper texture, visible pencil sketch strokes in garment folds, premium editorial masterpiece.

Food poster (3:4)

Japanese tonkotsu ramen promotional poster. Black ceramic bowl with rich creamy broth, soft-boiled egg halved, chashu pork belly slices, crisp nori, vibrant scallions. Dramatic steam curling upward against a near-black background. Vertical Japanese text "ラーメン" beside the bowl. Michelin-level food photography, award-winning chiaroscuro composition.

Athletic campaign (9:16)

Premium athletic brand campaign poster. Male sprinter exploding from starting blocks, chalk dust frozen mid-air, veins and sweat hyperreal. Matte black compression kit with subtle reflective strips. Deep charcoal gradient background with blazing amber motion trails. Large metallic headline "FORGE" top left, subhead "BREAK LIMITS" in sharp modern sans-serif woven into smoke. Cinematic sports photography masterpiece.

Fantasy key art (16:9)

Epic fantasy video game key art. Colossal crystal-armored serpent rising from a cracked desert arena, four heroes in dramatic combat poses below, storm clouds and lightning. Title text "ABYSS WARDEN" in bold metallic serif across the top sky. Cinematic Unreal Engine lighting, ultra-detailed VFX, AAA promotional quality.

Narrative still (3:2)

Astronaut floating in the ISS cupola, both hands wrapped around a warm mug, eyes on a tablet clipped to the thigh—not looking at Earth. Blue planet fills the curved windows behind, soft rim light on suit fabric. Quiet documentary NASA authenticity, natural film grain, intimate narrative moment, no text.

How MAI Image 2.5 Edit Works

Upload the source image, write what should change, generate. The model targets surgical edits: palette swaps, background replacement, layout restyles, art-direction shifts, and headline updates while keeping identity and composition stable across iterations.

Verified edit examples

Campaign localization (from an existing athletic ad)

Localize this campaign for a winter launch: replace the violet energy smoke with icy cyan and white frost particles, swap headline to "BEYOND LIMITS" in brushed silver sans-serif, shift background gradient to deep navy and glacier blue. Preserve the athlete's pose, outfit silhouette, and dynamic composition exactly.

Poster restyle (from a recipe layout)

Transform this recipe poster into a dark chocolate soufflé edition: hero dish becomes a rising chocolate soufflé in a copper ramekin, warm autumn palette, title text "DARK CHOCOLATE SOUFFLÉ" in refined serif. Keep the elegant step-by-step layout structure and cream-to-caramel background warmth.

Character re-scene (from stylized game art)

Place this tiny sorcerer character drifting above calm bioluminescent ocean waves at night, teal magic trail beneath the skiff, soft moon and stars above. Preserve the blue cloak, glowing yellow eyes, crystal staff, and stylized proportions exactly. Dreamlike roguelike key art, peaceful mood, no text.

Illustration to graphite

Convert this watercolor caricature into a polished graphite pencil portrait on white paper: retain the beanie, square glasses, craggy nose, and three-quarter angle exactly. Fine cross-hatching shading, gallery illustration quality, no color.

Cartoon to product photo

Turn this cartoon barbarian into a premium PVC collectible figure product photo: glossy painted statue on a circular display base, blister-pack style box blurred in background, studio softbox lighting. Preserve character design, axe, and color palette exactly.

Content vs instruction trap: Style and camera notes belong in the prompt. Literal customer-facing copy belongs in quotes so it renders as visible text, not ignored metadata.

Using the Two Models Together

Typical pipeline: generate a hero frame on MAI Image 2.5, then open MAI Image 2.5 Edit for localized fixes without re-prompting from scratch. Useful for seasonal palette swaps, market-specific backgrounds, headline localization, or turning a flat illustration into a photoreal deliverable.

For edit-only workflows, upload your own image or pick one from the Scenario library.

Use Cases

• Fashion and editorial: Magazine covers with mastheads, date lines, and column headlines in one generate pass.

• Food and beverage: Ramen, pastry, or menu posters with steam, chiaroscuro, and embedded type.

• Marketing: Sports and lifestyle campaigns with integrated headlines and motion effects.

• Games: Fantasy key art with title treatments; Edit to re-scene characters or convert stylized art to product shots.

• Film and narrative: Documentary-style stills with intentional story beats (subject not looking at the obvious focal point).

• Localization: Swap palette, season, and headline on an existing ad without rebuilding the layout.

Tips for Better Results

1. Write long, specific prompts. Name lighting, lens, layout zones, materials, and mood. One-line prompts rarely match pinned gallery quality.

2. Quote text that must render verbatim. Use "FORGE" and "BREAK LIMITS" in the prompt, not paraphrases.

3. Set aspect ratio explicitly for final delivery. Auto works for exploration; lock 9:16, 16:9, or 2:3 before handoff.

4. On Edit, state what to preserve. "Preserve pose, outfit silhouette, and layout structure" reduces drift.

5. One edit intent per instruction. Split a background swap and a headline change into two runs if results mix.

6. Use one reference per edit run. Describe products or secondary subjects in prose when dual uploads fail.

7. Use Image Count for type and layout checks. Generate four variants when composition is right but spelling is off.

8. Pair generator + editor for campaigns. One master generate, several localized Edit passes on distinct sources.

Known Limitations

• ~1 MP output cap. No native 2K or 4K on Scenario; GPT Image 2 may fit larger deliverables.

• Content moderation on Edit. Some action-heavy scene prompts may flag; rephrase toward environment and mood.

• No seed or output format controls. Scenario uses provider defaults.

• Plan access may apply. Access restriction level 25 (Generate) and 50 (Edit) on some workspaces.

• Cold start. First jobs may sit in warming-up while Fal provisions the endpoint. Submit one job at a time during heavy queues.

Open the models: MAI Image 2.5 · MAI Image 2.5 Edit