Adobe Certified Professional - Adobe Firefly cheat sheet

Generative AI models learn statistical patterns from large training datasets and produce new outputs (images, text, audio, video) that resemble the data distribution without copying specific training examples.

Adobe Firefly is a family of generative models trained on Adobe Stock and public-domain/licensed content, designed to be commercially safe for use in Creative Cloud workflows.

Firefly model families include image models (text-to-image, generative fill/expand), vector (text-to-vector), video (text-to-video, image-to-video), and audio/design models; model availability changes over time.

A prompt is the text instruction; generation parameters (aspect ratio, content type, style, seed, model version) control output variation and format.

Generative Fill and Generative Expand are inpainting/outpainting operations: they synthesize pixels inside a selection or beyond canvas edges using surrounding context plus a prompt.

Firefly outputs are probabilistic: identical prompts can yield different results; seeds and model versions help reproduce or vary outcomes.

Firefly is integrated across Adobe apps (Photoshop, Illustrator, Express, Premiere Pro, InDesign) and via Firefly web app and APIs, with generative credits governing usage.

Content Credentials (C2PA) attach provenance metadata to Firefly-generated assets, indicating AI generation and edit history.

Effective prompts specify subject, action, setting, lighting, style, camera/composition, and color palette; specificity reduces ambiguous outputs.

Prompt structure often follows: [subject] + [attributes/action] + [environment] + [lighting/mood] + [style/medium] + [composition/camera].

Negative prompts or exclusion phrasing can suppress unwanted elements; effectiveness varies by model version.

Aspect ratio and resolution are set before generation; changing ratio alters composition and may crop or extend content.

Style presets and reference images bias the model toward a visual look; strength sliders control how much the reference influences output.

Seeds allow reproducible generations when model version and parameters are unchanged; varying the seed explores variations.

Iterative refinement-generate, adjust prompt or parameters, regenerate-is standard; small prompt edits produce measurable output changes.

Text rendering in generated images is unreliable; legibility and spelling often require post-editing or vector/text tools.

Generative Fill requires a selection; the model synthesizes content inside the selection using the prompt and surrounding pixels.

Generative Expand extends the canvas using outpainting; the original image remains and new pixels are generated to fill extended areas.

Inpainting quality depends on selection size, feathering, and context; overly large or unfeathered selections can produce seams or inconsistent lighting.

Remove/erase tools use generative models to reconstruct background behind removed objects; results vary with background complexity.

Generative upscaling/enhancement increases resolution and detail; it cannot recover information absent from the source and may introduce artifacts.

Generative layers in Photoshop are non-destructive: prompts and settings can be revisited and regenerated without altering the base pixels.

Color, lighting, and perspective matching between generated and original content improves realism; mismatched shadows or tones reveal edits.

Iterative compositing-generating multiple variations and blending-is common; each generation consumes credits and adds to provenance history.

Text-to-video generates short clips from prompts; image-to-video animates a still image, and video-to-video restyles or transforms existing footage.

Video generation parameters include duration, aspect ratio, motion strength, camera movement, and frame rate; longer clips may be assembled from shorter generations.

Generative audio can produce sound effects, ambience, and music beds; speech generation (text-to-speech) is a separate capability with voice controls.

In design workflows, Firefly powers text-to-vector, generative recolor, and template population in Illustrator and Express.

Generative recolor applies palettes to vector artwork while preserving structure; results depend on how artwork is grouped and layered.

Firefly-generated assets can be placed into Premiere Pro, After Effects, and Express timelines for compositing with conventional media.

Generative credits are consumed per operation; credit costs vary by model and feature, and plans include monthly allocations.

Workflow integration matters: assets retain Content Credentials through Adobe apps, supporting downstream provenance tracking.

Adobe designs Firefly to be commercially safe: training data is Adobe Stock and licensed/public-domain content, and outputs are intended for commercial use under Adobe's terms.

Adobe's generative AI user guidelines prohibit generating illegal content, deepfakes intended to deceive, harassment, and content violating third-party rights.

Content Credentials are cryptographically signed provenance metadata based on the C2PA standard; they can record AI generation, edits, and tool history.

Content Credentials are optional to display but are attached by default in many Adobe workflows; they can be inspected via verify tools.

Users are responsible for ensuring prompts and outputs do not infringe trademarks, copyrights, or publicity rights; model output is not a guarantee of non-infringement.

Disclosure of AI-generated content may be required by law, platform policy, or contract; Content Credentials support but do not replace disclosure.

Firefly outputs may not be copyrightable in some jurisdictions because they lack human authorship; human creative contribution affects protectability.

Enterprise and API usage are governed by separate terms; credit allocations, data handling, and indemnification vary by plan.