Microsoft moved image editing up the leaderboard. Test the edit, not the rank

THE SIGNAL IN ONE SENTENCE

A leaderboard can tell you which door to open. It cannot tell you whether the room contains your furniture. Microsoft AI announced MAI Image 2.7 Preview on October 27. The company says the image-editing model debuted at number three on Arena Edit and number two on Artificial Analysis. The second table is especially useful because it publishes more than a single rank. On October 28, Artificial Analysis listed MAI Image 2.7 Preview at 1,193 Elo from 4,891 samples, with a 95 percent interval from 1,178 to 1,208. The table placed it behind GPT Image 2.5 Sunburst at 1,199 and ahead of GPT Image 2.5 Flare at 1,182. Those intervals overlap. So the honest reading is not that one system has permanently defeated another. It is that the preview sits in the current top group under this particular pairwise preference test. Artificial Analysis describes the evaluation as people comparing two edited images produced from the same source image and instruction, without the model names deciding the vote. The resulting Elo score reflects which output people preferred more often. That is meaningful evidence. It is not the same thing as a creative team approving a file for print, a product page, a film shot, a historical archive or a regulated campaign. Microsoft says 2.7 improved over MAI Image 2.6 across all 10 use cases and all seven editing actions evaluated by Artificial Analysis. It highlights enhancement and restoration, reasoning-based edits, scene and style changes, and text or symbol editing. The company says the model leads the next system by 24 Elo in scene and style. The examples of intended work are familiar: move a product into a new setting, change the lighting, alter one object without breaking the surrounding image, repair an older asset or replace text while keeping it visually integrated. That is the good news. The useful release is also narrower than the headline can sound. Microsoft calls the model a preview. Editing can be tried through MAI Playground, Arena and Artificial Analysis. Text-to-image generation is coming soon rather than available now. The Artificial Analysis table lists no API for 2.7, and Microsoft's announcement does not publish production pricing, latency, rate limits, licensing terms, a full model card or a separate safety evaluation. In other words, a studio can inspect the editing behavior today. It cannot yet infer a complete production contract. The gap between those two things is where an acceptance set earns its keep. Start with identity preservation. Use people, pets, products, rooms and landmarks that your team is permitted to test. Ask for a narrow change such as a jacket color, a background object or a lighting adjustment. Compare faces, body proportions, distinctive marks, packaging geometry and scene layout before and after. Do not score only whether the requested change appeared. Score what should not have changed. An image editor can satisfy the sentence and still quietly move an eye, invent a label, flatten a surface, alter a hand or replace the exact product with a plausible cousin. Those are not decorative errors when the subject is a real person or a catalog item. Next, test local edits. Mask one object or describe one region. Ask the model to change only that area. Measure spill into the background, shadows, reflections, depth of field and nearby objects. A red cup becoming blue is not a clean edit if the tablecloth, window light and person's sleeve also drift. Then test composition under pressure. Move a subject, add an object behind it, change the viewpoint or alter the environment. Check occlusion, perspective, contact shadows, reflections and physical support. An object should not hover merely because the prompt forgot to mention gravity. Text and symbols need their own lane. Use permitted mock labels, signs and interface elements with known spelling. Test short words, long words, repeated characters, punctuation, mixed case and the scripts your audience actually reads. Verify that the model changes only the requested text and keeps typography, alignment, contrast and neighboring marks stable. A pretty label with one wrong character is still wrong. Restoration also needs a truth boundary. Separate removing damage from inventing missing history. A model may clean scratches, balance tone and improve legibility. If it reconstructs a face, uniform, storefront or document detail that the source never preserved, label the result as a generated interpretation. Do not let a confident texture become archival evidence. Relighting should survive a physics check. Change time of day, mood or light direction. Inspect every shadow, reflection, highlight and translucent material. If the key light moves but the window reflection does not, the image may feel wrong before a viewer can explain why. Now repeat everything. Run each instruction several times with the same input and settings. Record how often the requested change lands, how often protected details drift and how many attempts a reviewer needs before approval. The cost of an editing model is not just a listed price per image. It is generations multiplied by retries, reviewer minutes, rejected outputs, file preparation and downstream correction. There is no published 2.7 API price to compare yet. Even when one arrives, a cheaper generation can be a more expensive approved asset if it requires five rounds and a retoucher. Build adversarial pairs. Put a protected logo near the edit area. Use a person with asymmetrical features. Include a mirrored surface, a transparent bottle, fine hair, patterned fabric, small print and a shadow crossing the mask. Try instructions that contain two changes but explicitly forbid a third. The goal is not to embarrass the model with a carnival trick. It is to find the point where your normal work stops being reliable. Keep provenance attached. Every test output should carry the source asset, permission basis, prompt, edit region, model name, model version, settings, timestamp, reviewer, decision and reason. Preserve the untouched original. If the product later exposes content credentials or other provenance metadata, verify that your export, resize, retouch and publishing chain does not strip it. Rights need a separate decision from quality. A model can produce an excellent edit that a team should not use. Confirm the rights to the source image, the authority to alter the subject, any restrictions on likenesses or cultural material, and the terms governing outputs. Do not use a leaderboard as a substitute for legal review. Create a stop list. Some jobs should remain human-led or require a second reviewer: identity-sensitive portraits, evidence, medical images, journalism, historical restoration, political material, children's images and regulated claims. The preview announcement does not establish suitability for those uses. Finally, decide before the demo what counts as a pass. Choose a weighted threshold for instruction success, protected-region fidelity, text accuracy, physical consistency, repeatability, provenance, review time and cost. Keep a few tests hidden from the people tuning prompts. Compare against the current human workflow and at least one alternative model. Then rerun the set whenever the preview changes. The version number matters because leaderboards move. Models are added, vote counts grow and the rank can change while the actual score remains within uncertainty. Your own asset mix can move even faster. A fashion studio, game team, museum and grocery catalog do not share one definition of a good edit. MAI Image 2.7 Preview has earned attention. It appears competitive in large human-preference evaluations, improves on Microsoft's prior version across the categories the company cites and is available for hands-on editing tests now. That is enough to put it in the trial queue. Promotion to the production queue needs a different receipt. Real assets. Protected regions. Repeat runs. Rights. Provenance. Reviewer time. Failure examples. A named person willing to approve the result. The leaderboard opens the door. Your acceptance set decides whether the model gets a key.