Generating or editing visual media solves a different problem from interpreting it. A visual-understanding model tries to describe evidence that already exists; an image-generation model creates new pixels in response to a request. A prompt that asks for a product illustration, a background replacement or an edited video sequence therefore raises questions about fidelity, consent, branding and disclosure that ordinary captioning does not. The computer-vision portion of AI-103 explicitly includes image and video generation as well as editing workflows.
The technical task is to pick a supported model and editing method, provide only approved source media, control the output and verify that the result still meets the user’s original requirements. An appealing rendering is not proof that product details, accessibility or usage rights were respected. Model and region availability change, especially for newer video capabilities, so production designs must distinguish generally available features from previews.
Begin with a visual specification that can be checked
Suppose a retailer needs an image of a blue backpack against a plain background. Specify the nonnegotiable details: product color, zipper count, logo placement, image dimensions, background style and allowed modifications. A model may produce a beautiful backpack while changing one functional feature. If the image will be used to sell a physical item, that inconsistency can mislead customers.
Build the request around the role of the media. A conceptual hero illustration tolerates imaginative content; a regulated product photograph or architectural plan often does not. Record whether the source is a photograph, illustration or synthetic composite. When exact fidelity matters, use controlled image editing on approved assets and compare the output against product-reference images, rather than asking for a complete re-creation from text.
Choose between creation, inpainting and broader edits
Text-to-image generation creates a new scene from a description. Reference-conditioned generation may use approved visual examples to guide style or composition. Inpainting targets a specified region with a mask while leaving other areas intended to remain unchanged. A prompt-driven edit may change large parts of an image and can be less predictable than a tightly masked operation.
Select the narrowest editing mode that satisfies the request. If the only requirement is removing a background object, a region mask reduces the risk of changing the product itself. Validate the unmasked regions after editing anyway; some models can alter colors, text or edges beyond the intended area. Retain the original and edit parameters where permission and retention policy allow, so a reviewer can reconstruct what changed.
Treat image prompts as specifications, not guarantees
Useful prompts define subject, composition, lighting, intended style, prohibited alterations and output constraints. They should not become long streams of contradictory adjectives. A prompt that asks for both an exact original logo and a “creative reinterpretation of all details” leaves the model guessing. Resolve those conflicts in the application UI or preflight validation.
For repeatable production tasks, version prompt templates and approved reference assets. Include a small test set with difficult product colors, embedded text and multiple objects. Models often struggle with exact textual rendering or small logos; use validated overlays or conventional graphics workflows for information that must be pixel-perfect. Generated text in an image should not be treated as a trustworthy source of labels, prices or compliance statements.
Plan video generation around temporal consistency
Video adds continuity requirements. An object should remain identifiable from frame to frame, motion should match the intended action, and an edit should not cause a person or product to change appearance unexpectedly. A short coherent clip is often preferable to a longer one that introduces impossible transitions. A video job may also have different duration limits, processing time and queueing behavior from image generation.
Define review criteria for motion, camera movement, temporal consistency, safety and brand representation. If the output will be published, inspect representative key frames and the entire motion sequence rather than judging only a generated thumbnail. Keep audio rights and synchronization separate when the workflow adds a soundtrack or spoken narration. Confirm the exact model capability and deployment region from current Microsoft documentation before promising production video editing on a particular service.
Suppose an insurance company needs a short instructional video showing how to photograph a damaged appliance. A single attractive frame is not enough: the appliance, label, camera perspective and sequence of actions must remain coherent across multiple seconds. Before generation, write a shot list that names the invariant objects and the actions in order. Each scene should have a clear acceptance criterion, such as “the serial number stays visible during the close-up,” rather than the vague requirement that the clip feel professional.
Review both the beginning and end of each generated segment. A hand may appear on the wrong side of an object, an equipment label may change text, or a safety step may be omitted. Short segments make it easier to correct one failure without regenerating the entire story, but cuts introduce another risk: the edited sequence can imply a physically impossible process. Keep the source specification and the actual frames available to reviewers so transitions can be audited rather than judged from memory.
Editing controls must be chosen for the actual correction. An inpainting mask may be sufficient to remove a prohibited logo from one still image, but changing a central object across a moving shot requires a tool that can preserve temporal coherence. Re-run content and brand checks after any meaningful edit; the revised output is a new artifact. Prompts can express intent, but the system should not assume that an image model guarantees pixel-accurate text or continuity.
The final publication path needs provenance. Store the licensed reference media, approved script, generated clips, human review record, moderation results and final export version. If the workflow includes a real employee’s likeness or voice, secure appropriate permission and avoid implying a generated demonstration documents an actual event. The engineering goal is repeatable, reviewed media production, not just a prompt that produces appealing frames.
Account for rights, consent and sensitive subjects
A user might supply a face photo, trademarked packaging or proprietary design. Application policy should define whether the submitted source is authorized for transformation and which edits are allowed. Do not assume a successful model call resolves licensing, privacy or consent issues. Record provenance of source assets and apply retention controls for uploaded media.
Public-facing content may need disclosure that a scene was generated or materially altered. Avoid creating visuals that imply an individual endorsed a product or that a fictional event actually occurred. Review local legal and organizational requirements for likeness, brand usage and synthetic media. Responsible AI is an operating process spanning input approval, generation, review and distribution—not merely a content filter on the final image.
Use filters and review to manage visual risk
Safety services may detect disallowed or inappropriate content in inputs and generated output. Define what the application does after a detection: refuse, offer a narrower permitted edit, or route to human review. An automated block may be necessary for some categories; in a legitimate moderation or accessibility workflow, annotation and escalation may be more appropriate. Test false positives to avoid turning a useful system into one that rejects ordinary tasks.
Brand checks are similarly layered. A classifier may flag a logo that should not be present, but a human reviewer may still need to confirm product fidelity and disclosure. Watermarking or provenance metadata can help indicate synthetic origin where supported, yet they do not prove an image depicts reality. The underlying source and approval records remain important for auditability.
Measure visual quality with targeted comparisons
Create a scoring checklist tied to each request: required object present, prohibited object absent, accurate color, logo fidelity, legible text, composition fit, unintended edit leakage and safety compliance. For video, add continuity, motion plausibility and audio alignment where relevant. A global aesthetic score should not override a failure in a mandatory attribute.
Evaluate multiple outputs for difficult prompts rather than assuming one successful example is representative. Use the same input and criteria when comparing models or prompt changes. Track revisions and approved assets alongside output version, model and settings. This lets a team understand whether a regression came from a new deployment or simply from a different reference image.
Design the workflow for cost and asynchronous jobs
Visual generation may be more expensive and slower than ordinary text inference. A web request should not remain open indefinitely while waiting for a long media job. Use a job identifier, status polling or notifications where the chosen service supports them, and make retry behavior idempotent. A client that times out after submitting a costly video request should check job state before submitting it again.
Store outputs in an access-controlled location with an expiry and an approval state. Distinguish “render completed” from “asset approved for publication.” If a user rejects an output, retain only what is needed for revision and operational logging. Do not keep every rejected image forever simply because object storage is convenient.
Practice safe visual edits for AI-103
For an AI-103 exercise, use a legally approved sample product image and request three changes: a background replacement, a mask-limited object removal and a short promotional motion concept. Compare each result with the original and note any unexpected modifications. Add a test where the prompt conflicts with a locked product attribute and verify that the application rejects or escalates the change.
Finally, inspect the pipeline from request to output: supported deployment, source rights, mask and prompt, moderation, asynchronous job control, quality review, accessibility description and publication approval. A robust media workflow can explain not just how visual content was generated, but why the final asset is appropriate and faithful to its intended use.