Image-generation models convert prompts and other supported inputs into new visual content. They are useful for concept sketches, educational illustrations and other creative workflows, but the result is not a photograph of a real event or automatically accurate technical documentation. Microsoft AI-901 expects learners to recognize generative image capabilities and understand how to create visual outputs through supported Foundry models. A basic solution also needs a way to review the generated content before it reaches people.
Consider a public science museum designing an illustration for a children’s exhibition about renewable energy. The museum wants a friendly image showing wind turbines, solar panels and a city landscape. Generative AI can speed up concept development. It can also accidentally depict a physically impossible turbine, include misleading safety details or create visual elements resembling someone’s protected artwork. The team needs a creative brief and an editorial review, not just an attractive first result.
Identify the actual generation goal
Specify whether the output is decorative, explanatory or intended as evidence. A decorative banner can tolerate creative variation. A technical diagram showing how to wire a solar array requires expert accuracy and should not be trusted merely because its labels look convincing. A fictional illustration should be labeled appropriately if viewers could mistake it for a real location or person.
If a museum needs to analyze an existing artifact photograph, image generation is not the correct first operation. Use computer vision or content extraction for interpreting the source, then create a separate illustration only when appropriate. The visual understanding explains this distinction through evidence-based scenarios.
Write a brief with constraints and an audience
A useful generation prompt names subject, visual style, composition, audience and any critical exclusions. For example, ask for a friendly, stylized cityscape with clear wind and solar energy elements, no text labels and no real-person likenesses. Avoid requesting nonexistent technical certifications or claiming that a generated design has been professionally reviewed. The system’s approval criteria should reflect the intended use, not merely the prompt’s adjectives.
For children, test whether the imagery is understandable at the relevant age and whether it represents different people respectfully. For web delivery, consider required aspect ratio, legibility at mobile size, alternative text and compatibility with the museum’s design system. Output dimensions and other generation options depend on the model and deployment; verify the current service rather than copying unsupported configuration parameters.
Choose a supported model and deployment
A text-only chat model may describe an image prompt without actually producing an image. Choose a Foundry model that explicitly supports visual generation and check its available input types, policy restrictions, regions and output handling. A model card’s examples are not proof that the output will fit every commercial or educational use. Confirm acceptable usage terms and whether the deployment is supported in the intended environment.
The AI-901 objectives lists creating new visual outputs with generative models. The Foundry deployment tutorial gives the resource-selection context. For exam purposes, distinguish the model’s capability from the deployment’s authorization and the application’s final editorial approval.
Evaluate visual correctness and misleading details
Build a simple review rubric: requested subjects present, unnecessary objects absent, no inappropriate or misleading content, visual coherence, reasonable accessibility and compliance with the creative brief. For the renewable-energy image, a turbine’s blades and solar panel layout should not imply a false engineering mechanism. If the illustration communicates a real-world scientific fact, a subject expert should validate it.
Pay attention to generated text inside images. Some models may produce distorted spelling or plausible-looking labels that are wrong. Rather than relying on embedded AI lettering for a precise title or statistic, add approved typography in a separate design process. Check output under different crops and sizes so important details are not cut off by a mobile layout.
Address rights, identity and provenance
Generated visuals can raise licensing, attribution and brand concerns. Review the applicable model terms and organizational policy before publication. Avoid asking a model to reproduce an identifiable living artist’s specific work or simulate a real person’s endorsement. Do not present a generated scene as documentary evidence. A clear provenance note may be appropriate when the image could otherwise be misunderstood.
A museum may combine human-created drawings, stock photography and generated elements. Track source and license for each component, and record edits and approvals. This is not a claim that all AI-generated content is legally identical; rights and restrictions depend on the jurisdiction, source and product terms. The responsible choice is to verify before commercial release rather than assume a paid model subscription grants unrestricted rights.
Protect privacy in image editing and generation
If an application accepts a user’s image as conditioning input, identify whether that image contains faces, addresses, badges or other personal information. Do not upload sensitive member photographs solely to create a decorative banner. Use approved synthetic examples during development. Decide how long uploaded images and generated outputs are retained, which staff may access them and whether the service’s data-processing terms match organizational requirements.
For a classroom demonstration, use fictional subjects and nonprivate images. If a generated image includes a realistic person, do not infer the existence, identity or consent of a real individual. The responsible-AI scenarios connect privacy and transparency to creative use cases without treating a content filter as a complete risk assessment.
Test revisions rather than selecting by intuition
Create three versions of the same museum prompt: a vague sentence, a detailed scene brief and a brief with explicit exclusions. Generate permitted outputs and evaluate each against the same rubric. Record whether extra instructions improved the requested composition or introduced an unforeseen trade-off. A favorite image from one run is not evidence of consistent behavior. Multiple samples can reveal when a model tends to omit or distort specific elements.
Measure the cost and time needed to produce a publishable image, including human review and correction. If a model generates ten drafts but none meets the factual brief, the apparent speed advantage is misleading. For educational use, a manually drawn diagram can sometimes be the more reliable tool, especially when precise geometry or labeling is required.
Deliver a responsible final artifact
The museum’s workflow should include a prompt record, selected model/deployment, output candidate, review notes, rights check, accessibility text and final human approval. Maintain a distinct version of the published visual and avoid silently replacing it after a lesson has been released. If a generated illustration contains a factual error, update the caption or regenerate it and record the change.
AI-901’s foundational lesson is simple but important: generative image capability creates new content; visual understanding interprets existing evidence. Choose between them deliberately, verify actual model support and test outputs against audience, safety, rights and accuracy requirements before presenting them as finished work.