Photoshop's generative tools have been in the hands of working retouchers long enough now for the honeymoon to be over. The demonstrations still look magical. The billable reality is narrower and more interesting: there is a set of jobs where Generative Fill saves an hour, a set where it saves nothing, and a set where using it is a professional risk. Knowing which is which is the actual skill.

What the tools do

Three features cover most of the work. Generative Fill replaces a selection with new pixels, guided by an optional text prompt and by the surrounding image. Generative Expand extends the canvas and invents what was beyond the frame. The Remove tool paints out an object and reconstructs what was behind it, which is the same idea with a brush instead of a selection.

Two things about the implementation matter for real work. The result arrives as a new layer with a mask rather than being burned into your image, so it is editable and reversible. And each generation returns several variations, so the first result is a draft, not a verdict.

Where it genuinely earns its place

Reformatting for a different aspect ratio. This is the strongest use by a distance. A landscape photograph that has to run as a vertical banner, a square social crop that now needs to be 16:9, a hero image with nowhere to put the headline. Generative Expand fills the new space with background rather than sky-blue guesswork, and because the invented area is background rather than subject, mistakes are cheap.

Removing distractions from a background. A bin, a stray cable, a reflection in a window, a person walking through the far edge of the frame. The reconstruction only has to be plausible, not correct, and nobody will ever compare it to reality.

Filling the gaps after a geometric correction. Straighten a horizon or fix converging verticals on a building and you are left with wedges of empty canvas at the corners. Generative Fill closes them in seconds.

Rebuilding simple, repeating texture. Tarmac, grass, sand, plaster, water, foliage. Content-aware fill was already decent at this; the generative version is better and needs less cleaning up.

Concept comps and internal layouts. Placing a rough visual in front of a client to agree the composition before the shoot is booked. Say clearly that it is a comp, and reshoot or license the real image for production.

Where it does not help, and where it is a risk

Faces, hands and people at any size. Generated people fail in ways that a viewer notices without being able to say why: eyes that do not match, teeth that go wrong, a hand with an extra knuckle. If a person is a recognisable part of the frame, do not generate them.

Anything the client sells. Do not use generative fill on the product, its label, its logo, its stitching or its packaging text. A regenerated bottle cap is a different bottle cap. This is not a quality judgement, it is a factual one: you would be showing a customer a product that does not exist in that form, and in regulated categories that is a legal problem rather than a design one.

Text of any kind. Generated lettering is unreliable and, when it does come out clean, it is not typeset. Set type in Photoshop or Illustrator.

Precise geometry and repeating structure. Window grids, tiled floors, brickwork, machinery, wheel spokes. Perspective drifts and repeats go out of step. Cloning or a manual patch is usually faster than fighting it.

Journalism, documentary and evidence. Removing a distraction from a news photograph changes the record. Most newsrooms prohibit it outright. The same principle applies to before-and-after images, insurance photography, property listings and anything a person will make a decision from.

Large fills in a high-resolution file. This limitation surprises people. The generated area is rendered at a modest resolution and scaled to fit your selection, so on a 60-megapixel file a large fill can come back noticeably softer and cleaner than everything around it. Keep generative selections small relative to the image, and expect to add grain and sharpening to make them sit.

Rights, credits and telling the client

Adobe states that the Firefly models behind these features are trained on Adobe Stock content along with openly licensed and public domain material, and it offers an indemnification arrangement to enterprise customers for certain Firefly-powered outputs. The terms are Adobe's and they change, so read the current ones for your plan rather than relying on what a colleague remembers. The same goes for generative credits: usage is metered, the allowance depends on your subscription, and behaviour when you run out has been revised more than once.

Content Credentials are attached to content produced with these features, which is a provenance record travelling with the file. Treat that as a reason to be straightforward with clients rather than a problem to route around. In practice, one line in the handover note is enough: which images contain generated areas and what was changed. Studios that do this have easy conversations later. Studios that do not, occasionally have very bad ones.

If you work with a public sector client, a bank or a listed company in Colombo, expect this to appear in contracts before long. Having an answer ready is a small competitive advantage.

How to get better results

Select beyond the edge. Include a margin of the surrounding area in the selection so the model has context to match. A tight selection produces a visible seam.

Describe what should be there, not what you want removed. "Wet tarmac in shade" works. "Remove the car" does not, because the model is filling, not deleting.

Keep prompts short and material. Surface, light and direction. Long descriptive prompts on a small fill mostly add noise.

Generate in several small passes. Three modest fills beat one large one, and each can be masked independently if only part of it is good.

Use the variations properly. Take the best parts of two results and mask between them. Nobody has to know that the final background is a composite of three generations.

Finish by hand. Match grain, match noise, match colour temperature, then sharpen the generated area to the same level as its surroundings. This last five minutes is what separates work that looks retouched from work that looks generated.

A checklist before the file goes out

  • View at 100 per cent across every generated area, not fitted to screen.
  • Check edges where generated meets original for softness, halo or a change in grain.
  • Look for repeated detail: the same leaf, the same paving slab, the same cloud twice.
  • Confirm nothing was invented on the product, on a face, on text or on a logo.
  • Keep the generative layers live in the working file so a revision does not mean starting again.
  • Note in the handover what was generated.

Used with judgement, these tools remove the dull hour from a retouching job: the corners, the crops, the cable in the background. They do not remove the need to know how selections, masks and colour work, and a retoucher who cannot finish a generated area by hand produces worse work with them than without.

Our Photoshop and generative AI course covers exactly this boundary, on real files, including the failure cases above so you can see them happen rather than take our word for it.