Adobe Certified Professional - Adobe Firefly common mistakes
Adobe Certified Professional - Adobe Firefly · Understand the traps and how to avoid them.
Treating Firefly as a model that stitches together or cites existing images, and describing outputs as remixes of training images.
Candidates carry over mental models from search-based or collage tools and from popular discussion about how generative models work, so they misdescribe the generation process.
What to do instead
Learn the Firefly fundamentals framing: outputs are newly generated from learned patterns in response to a prompt, not assembled from a library of retrieved images. Be able to state what a diffusion-style generation pipeline does at a high level without claiming retrieval.
Assuming Firefly has only one model or that all generation happens in a single place, then answering questions about model choice, resolution, or where features live incorrectly.
Candidates practice with one entry point (often the web app) and never explore the standalone Firefly site versus Firefly features embedded in Creative Cloud apps, or the different model options offered for image and video.
What to do instead
Map the surface area: standalone Firefly web app, Firefly inside Photoshop/Illustrator/Express, and partner-model access. Note which controls (model selection, aspect ratio, content type, effects) exist in each context and how outputs route into editing workflows.
Writing short, vague prompts like 'a nice picture of a dog' and expecting usable results, then blaming the tool.
Candidates under-specify because they assume the model will infer intent, and they do not iterate. Exam items about prompt crafting reward specificity about subject, style, lighting, composition, and medium.
What to do instead
Build prompts deliberately: subject, action, setting, style/medium, lighting, camera or composition cues, and mood. Practise revising a weak prompt into a strong one and predicting which change alters which attribute.
Confusing negative prompting, style/reference influence, and structural controls, or assuming a text prompt alone can enforce exact composition or layout.
These controls look similar in a UI and candidates memorise names without understanding what each one actually constrains, so they pick the wrong tool for a stated goal.
What to do instead
For each control, learn the question it answers: what to exclude, what visual style to borrow, and what structure or composition to hold. Match the control to the goal in the scenario rather than to the feature name.
Expecting Generative Fill or Generative Expand to behave like a normal clone/heal or crop, and mispredicting what happens to the surrounding pixels.
Candidates treat generative editing as a variant of retouching, so they miss that selections define the region to be regenerated, that prompts steer the fill, and that expand extends the canvas with newly generated content.
What to do instead
Practise the selection-first workflow: make a selection, describe the intended content, generate, then evaluate variants. Learn that Generative Expand requires extending the canvas before generating, and that results arrive as selectable variations on separate layers so the original is preserved.
Ignoring non-destructive workflow details such as layers, masks, and which object receives the generated result, then answering editing questions as if changes were flattened and permanent.
Candidates focus on the generation itself and not on how Firefly integrates with the host app's document model, so they cannot reason about reversibility or compositing.
What to do instead
Study where generated output lands (new layer, generative layer, masked region), how to revisit and regenerate, and how to keep the original intact. Be able to describe how to composite generated content with existing imagery.
Assuming generative video works exactly like text-to-image with a single prompt, or that audio and design features share the same controls and limits.
Candidates generalise from image generation and skip the specifics of video (duration, motion, image-to-video, frame or camera behaviour) and of audio and design workflows.
What to do instead
Learn each media type's inputs and constraints separately: what video generation accepts (text, reference image, motion or camera direction) and typical clip length, plus how audio and template-based design generation differ from image prompting.
Treating Content Credentials as a watermark or a setting that can be switched off to make output untraceable, or believing they prove authorship.
Candidates conflate provenance metadata with visible watermarks and with legal authorship, so they misstate what the credentials attach to and what they assert.
What to do instead
Learn that Content Credentials are cryptographically signed provenance metadata (C2PA-style) attached to the asset describing how it was made and edited, that they travel with the file, and that they indicate provenance rather than ownership or copyright.
Claiming that Firefly output is automatically free of third-party IP concerns in every use, or conversely that any generated image is unusable commercially.
Candidates collapse a nuanced licensing position into an absolute rule, often because they heard a slogan about commercially safe training data without learning the conditions and exceptions.
What to do instead
Learn the intended-use framing: Firefly is designed for commercial use with training on licensed and public-domain content, but usage is governed by the applicable terms and by the customer's context, and some features or partner models carry different conditions. Answer based on the stated terms, not absolutes.
Skipping disclosure and human-review obligations, or generating depictions of identifiable people, brands, or sensitive content without considering policy limits.
Candidates focus on technical capability and treat responsible-use rules as boilerplate, so they miss scenario questions about transparency, review, and restricted content.
What to do instead
Memorise the responsible-use posture: disclose AI-generated content where appropriate, keep a human in the loop to review outputs for accuracy and bias, respect limits on depicting real people and protected marks, and follow the content policies. Apply these to scenario stems rather than reciting them.