Distinguish visual questions from text recognition

An image-understanding model can describe an office, identify a chart’s general topic or answer simple visual questions. An OCR or document-layout workflow instead tries to identify exact characters, reading order, table relationships or fields. A model that says a photo contains an invoice has not established the invoice total. Decide whether the user needs a general description or a value that must be copied accurately into a business record.

Choose three permitted images: a simple scene, a chart with a clearly labeled axis and a receipt containing a small amount. Ask a broad question and an exact-value question for each. Record which output is supported by visible pixels and where the model guesses. For the receipt, compare the returned amount with the original text using an independent check. This helps candidates learn when structured information extraction is preferable to a conversational image description.

Place a product box with a serial number beside an unrelated delivery slip in one photograph. Ask a visual model to describe the scene and then extract the serial number. It may correctly identify a shipment but attach a number from the delivery slip to the product. A specialist OCR or structured extractor with source-region evidence makes that confusion easier to detect. Keep the crop and source ID with any accepted value, and refuse to claim an exact serial number when the text is blurred.

Create useful alternative text without overclaiming

Alt text exists to communicate an image’s function or relevant information to people who cannot see it. A decorative flourish may need empty alt text; an informative process diagram needs the essential relationships conveyed in concise language. A model can draft a description but may miss a critical label or invent a visual element. Human context determines which details matter. An alt text generator should not assume identity, intent or sensitive attributes from a person’s appearance.

For a simple accessibility exercise, describe a chart with two lines and one meaningful change. Ask the model for concise alt text, then compare it with the actual data and the purpose of the surrounding article. A description that merely says “a chart” is too vague, while a paragraph of speculative business interpretation is unhelpful. Test an image where the numeric label is unreadable; the model should acknowledge uncertainty rather than fabricate a value. Measure how often an authorized reviewer edits the proposed text.

Compare two proposed descriptions of an error-rate chart. One says “line graph showing improvement,” while the other identifies which axis measures percentage, the approximate range and the visual trend. If the chart uses nonstandard colors or a logarithmic scale, a simplistic description can mislead. An authorized subject reviewer should decide the minimal essential detail based on surrounding prose. Never infer causation merely because two plotted values move together. This is an accessible-content task with independent factual constraints.

Select image generation according to the intended artifact

Image generation creates pixels from instructions. Editing may change a source image while attempting to preserve specified regions. These are different operations and model capabilities, and not every Foundry deployment supports masks, reference images, transparency, the same resolution settings or all content types. Verify the exact model and API support before claiming an editing workflow is available. A catalog entry or marketing demonstration is not the same as an enabled feature in the learner’s tenant.

Write an asset brief with aspect ratio, audience, allowed visual style, required objects, forbidden content and the delivery channel. For example, a generic classroom illustration may require two desks, a whiteboard and no real company branding. A successful generated image should match the brief without adding misleading text. Compare a first prompt with one clearly constrained revision while holding other settings stable. Keep prompt and model versions so results can be reviewed, but avoid promising identical outputs from every run.

Evaluate edits by what must remain unchanged

For an edit test, use a source picture the learner has rights to modify. Ask for a background change while preserving an object’s geometry, label and visible warnings. If the selected model supports masks or protected regions, document how those boundaries were supplied. Compare the edited result with the original in the protected areas. An aesthetically attractive image that changes a required product label is not a successful edit.

Add a deliberately difficult test containing small printed text. Models can produce believable but incorrect labels; for an instructional image the wrong warning may be more harmful than visual artifacts. Record whether the edit was truly supported by the chosen model and deployment. If the model cannot guarantee preservation of legally important text, use conventional editing or a person rather than treating “make no changes elsewhere” as technical proof.

Use a safe fictional cereal package with a warning label. Ask for a background-only edit, then compare the warning text, object proportions and label placement to the original source. If the model changes a letter while preserving the background, the edit fails a protected-content criterion even if a human prefers the new aesthetics. Document whether the deployment supports a mask or only a text instruction, because those are not equivalent mechanisms. An unsupported preservation requirement should trigger a different workflow rather than repeated unbounded prompting.

Apply safety and intellectual-property gates

Before generating or editing, confirm that the organization has rights to reference images, likenesses, logos and characters. A prompt does not grant a third-party license or a person’s consent. Apply the applicable platform content filters and the organization’s publication policy. Input moderation alone cannot verify that a new output is factually accurate, brand compliant or legally distributable. Keep rejected outputs out of public feeds and retain only appropriate sanitized review metadata.

Use a simple checklist after generation: required content present, forbidden content absent, real-world claims accurate, reference permissions documented and user-facing accessibility supplied. If a watermark or disclosure is required for the channel, apply the relevant supported policy; a watermark does not prove factual correctness. A human reviewer should approve public high-impact graphics where the image conveys instructions, product details or an endorsement. Record the accepted exact file version to prevent later unreviewed edits being distributed.

Know when an image is insufficient evidence

A photograph may show someone near a control switch without proving that a safety procedure was followed. A screenshot can display a number without verifying that the dashboard was current. A generated illustration can depict an assembly that does not exist. Separate observed visual features from inferred actions and externally verifiable claims. Do not reward an answer for confidently describing what the image cannot establish.

Create two images of a similar scene with one meaningful difference, such as a warning label visible in one but obscured in the other. Ask the model a safety-relevant question. The correct response should distinguish the observed warning from the unreadable one, not fill both answers from likely context. Link the resulting note to the exact image version or crop used. This evidence discipline is a recurring AI-901 skill, not merely a special requirement for professional computer vision.

Test model updates with an independent visual sample

A small visual evaluation set should cover common scenes, dense charts, small text, unfamiliar objects and cases where the appropriate answer is uncertainty. Score caption relevance, source support, accessible-description usefulness and output-policy issues separately. For generated assets, compare object count, reference fidelity and forbidden additions; for edits, explicitly assess preservation boundaries. Average aesthetic preference cannot offset a factual diagram error.

Record model version, selected region, relevant generation parameters, exact input permissions and reviewer outcomes. When a new model becomes available, rerun a stable sample rather than assuming a larger model or newer preview automatically performs better. Some API controls may change across versions. Keep a fallback that does not expose private reference media or promise unsupported operations. This converts an image demo into a repeatable learning exercise.

Separate severe defects from cosmetic differences in your evaluation notes. A missing warning symbol, unauthorized logo or altered technical arrow may block approval. Minor lighting variation in a decorative image may be acceptable. When comparing two models, use the same source rights, prompts and accepted output dimensions, and disclose which controls were unsupported. A model producing eight attractive candidates may cost more than one producing a single approved asset. Counting returned pictures is not the same as measuring success.

Finish with a safe distribution decision

Combine the three tasks into a short decision report. For scene description, explain which visible objects justify the answer; for OCR-like extraction, identify which value requires an independent check; for generated art, document which requirements and usage rights were approved. If an image will be publicly shown, keep a review stage before release. Delete temporary reference files and generated candidates according to the test environment’s retention policy.

Microsoft’s AI-901 study guide emphasizes recognition and practical implementation of basic vision and image generation capabilities, not a promise of universal model fidelity. A defensible learner can choose the right visual operation, explain a model’s limits, apply a safety check and identify when a specialist extractor or human must take over.