Adobe Firefly models explained: the creative AI family and the Firefly app

Firefly is both a model family and a creative workspace

Adobe Firefly has two related meanings. First, it is Adobe’s family of generative AI models for creative tasks. Adobe says these models power generative features in its applications as well as its standalone Firefly experience. A model is the system that produces or transforms the media.

Second, Firefly is the app or platform around that capability. Think of the app as the place where you select a task, enter a prompt, add a reference, review variations and continue editing. It provides the canvas, controls and workflow rather than being a single model.

That naming overlap is the source of much confusion. A result made in the Firefly app may have used an Adobe model, but the app can also present a partner model for certain tasks. Check the selected model and workflow, not the Firefly name alone.

What Adobe’s own Firefly models are meant to create

Adobe describes its own Firefly models as a creative set spanning images, video, audio and vectors. Those categories describe the kind of material a model can help generate or change, rather than four interchangeable settings. An image task is useful for a visual concept or an edited composition. A vector task is aimed at editable graphic shapes. Video and audio tasks address different media and controls. The option available to you will depend on the feature you are using.

An Adobe Firefly image model is therefore best understood as a generative visual system, not a photo library search box. You provide an instruction in words and may be able to add a reference or make choices about the look of the output. The system produces new visual candidates based on those inputs. You can then judge the candidates and revise the instructions or edit the result. Adobe’s public materials describe the creative workflows, but they do not give readers a complete technical blueprint for each model’s internal architecture. It is better not to invent one.

Model choices can change. For a live project, use the label in the specific tool as the source of truth.

How an image-generation workflow works in practice

A typical image workflow begins with a creative brief, not a magic sentence. You decide the subject, the purpose of the image and the non-negotiable details. Then you choose an image-generation or image-editing task, select an available model, provide a prompt and, where the tool supports it, add an existing image as a reference. The chosen model and tool determine which controls are offered, so the same words may not produce the same type of output in every workflow.

Here is a hypothetical example. A community museum needs an early concept for a poster about a night exhibition. A designer might ask for a moody entrance scene, specify a vertical composition, leave clear space for later copy and use the first output only as a direction to discuss. If the doorway feels too modern or the lighting is too bright, the designer changes the request or edits the candidate rather than assuming the first generation is final. This is hypothetical, not a claim of personal use or a tested result.

The useful human work happens around the generation. Compare it with the brief, check text and objects, and note the selected model if the direction must be reproduced. A generated image is not evidence that a scene, object or written claim is accurate.

Why a Firefly model menu can include non-Adobe models

Firefly is also a multi-model creative environment. Adobe says that some Firefly workflows can offer choices from partner providers alongside Adobe’s own models. A partner model is not transformed into an Adobe Firefly model because you accessed it inside the Firefly app. It remains a model developed by that partner, even though Adobe supplies the surrounding workflow.

Models can differ in controls, creative behaviour and suitability. Adobe says people should be told when a partner model is used and are responsible for deciding whether it fits their project. Read the model name and any notice rather than treating all outputs as equivalent.

Partner availability can vary by region and change over time; Adobe also records retirements. Check the picker in the exact feature you intend to use, then record the choice for important work.

Commercial-safety claims still need human review

Adobe says its own Firefly models are trained on licensed material, such as Adobe Stock, and public-domain material where copyright has expired. It also says it does not train its Firefly models on customer content. These are Adobe’s published commitments about native Firefly models, not a substitute for reviewing every result.

For partner choices, Adobe says customer data and content in its apps are not used to train either Adobe or partner models in these integrations. It also says the creator must decide whether a partner model is suitable, including for commercial use. Do not carry a claim about Adobe’s own family over to every model in the app.

Before publishing or delivering an asset, check it for visual errors, unwanted reference details, identifiable people, brand marks and misleading text. Content Credentials can provide creation context, but not replace quality control, fact-checking or a rights decision.

FAQs about Adobe Firefly models

These answers are general because model choices, controls and availability can change between features and regions.

For a time-sensitive project, document the model name displayed in the exact Firefly workflow.