Photoshop Generative Fill: what it does well, and where it fails

Photoshop's Generative Fill adds, removes or replaces content in a selected area from a short text prompt, then puts the result on a new layer with a few variations to choose from. Leave the prompt empty and it fills the selection from the surrounding image. Generative Expand does the same beyond the edge of the canvas when you drag the Crop tool outwards.

These tools are very good at extending a background, clearing distractions, building scenes around real artwork and speeding up retouching. They are weak at hands, text, Arabic letters, exact brand colours and large areas at print resolution.

Everything about the tools below comes from Adobe's help pages as I checked them in October 2026, with each page's date. Then: where they fit in real design work, a careful layer-based workflow, credits, Content Credentials, and what Adobe's own terms say about commercial use. If you are new to the toolbar, start with my Photoshop tools guide. This article is part of my practical guide to AI in graphic design.

What Generative Fill and Generative Expand are

Adobe's Generative Fill page (updated 15 September 2026) describes five steps:

1. Make a selection with any selection tool, such as the Selection Brush or the Lasso.
2. Choose Generative Fill in the Contextual Task Bar, or Edit > Generative Fill.
3. Pick a model. Adobe's own: Firefly Fill & Expand, Firefly Image 5 and Firefly Image 1. Partner models: FLUX.2 pro and FLUX Kontext pro from Black Forest Labs, and Google's Gemini 3.1 (with Nano Banana 2) and Gemini 3 (with Nano Banana Pro).
4. Type a prompt, or leave it blank to fill the selection from the pixels around it.
5. Select Generate and choose a variation in the Properties panel. Each result arrives as a new generative layer with its own mask.

The same page notes that Google is retiring Gemini 2.5 (with Nano Banana) from 2 October 2026. Model lists change every few months, which is why this article carries dates.

Generative Expand works through the Crop tool: drag the handles beyond the image, choose Generative Expand in the Contextual Task Bar, and leave the prompt blank to continue the scene. Its help page (February 2026) lists Adobe's own Firefly models for it.

Adobe's FAQ adds three practical points: the features need an internet connection, generated layers need a layered file such as PSD to survive a save, and prompts are accepted in more than 100 languages.

The related tools in one table

Where it helps in real design work

Extending a background

The most common job: a horizontal photo has to become a 9:16 story, or a portrait shot a wide web banner. Generative Expand continues the sky, the wall or the table so the subject keeps its place. Check the new area for repeated textures and odd seams, and keep the product itself outside the generated part. I cover the sizes each platform needs in social media post design with AI.

Removing objects

A cable across a shopfront, a passer-by, a parked car, a stray sign. For small things, the Remove tool is quicker, and its Create new layer option keeps the photo untouched. For larger areas, select generously and run Generative Fill with an empty prompt.

Mockup scenes around real artwork

Generative Fill can build the world around a design: a café counter, a hand, a shelf. The rule is that the design itself stays real. Place the artwork as a Smart Object, select only the area around it, and generate the setting. If you place a product photo into a new background, Harmonize can match its light and colour, but it works on pixel layers only, so run it on a rasterised copy. Never ask the tool to draw your logo or your label.

Retouching

Dust, crumbs and reflections on packaging photography, a scratch on a product, a crease in a backdrop. Small areas give the best results, and each fix sits on a layer you can switch off. With photos of real people, keep the edits honest: clean a background, but do not change a face or a body in a way that misleads.

Its limits, and what to check

Hands and small details

Hands, fingers and small objects held in a hand are a well-known weak spot of image models. If a fill touches a hand, look at it at actual pixels before anyone else does.

Text and Arabic letters

Generated areas often contain letter-like marks, and Arabic comes out worse than Latin, for reasons I explain in Arabic text in AI images. Adobe's own guidelines say quality may be lower for languages it does not explicitly support. Remove any fake letters, then set real text with the Type tool, using World-Ready Layout for Arabic.

Brand colours

The model does not know your hex values, so generated areas drift towards whatever looks plausible. Generate in neutral terms, then bring the colour back with adjustment layers clipped to the generative layer, and check the result with the Color Sampler against the brand guide.

Resolution

Adobe's FAQ says the Firefly model returns 1024 × 1024 pixel output, so filling a large area of a big file stretches it and the result looks soft. Work in smaller pieces, use Enhance Detail where it is offered, and judge the result at the final print or screen size. Partner models offer larger outputs at a higher credit cost.

Truth

A product, a label, a medicine box, a real place: these have to stay as they are. Generative Fill can invent a convincing detail that was never there, which is the last thing a customer should find on a pack.

A careful workflow on a separate layer

1. Protect the original. Keep the source photo locked at the bottom of the stack, or work on a copy.
2. Select tightly, with a little overlap into the surrounding pixels so the new content blends.
3. Describe what should be there, in plain words, or leave the prompt empty to clear an area.
4. Compare the variations, and use Generate Similar when one is nearly right.
5. Refine the mask of the generative layer to soften the join.
6. Correct colour and light with clipped adjustment layers, not by regenerating again and again.
7. Check at actual pixels and at final size, looking for repeated textures, fake letters, extra fingers and warped edges.
8. Name each generative layer with the model you used.
9. Save as PSD so the generative layers survive.
10. Turn on Content Credentials (Window > Content Credentials (Beta)) before you export.

Credits and Content Credentials

Adobe's generative credits FAQ (updated 30 September 2026) says standard features such as Generative Fill with Firefly models use one credit per generation. Partner models count as premium features and use more, depending on the model and the output size. Some plans include unlimited standard generations, and credits reset each month without rolling over. Adobe's partner-model page lists Generative Fill with Gemini 3.1 at 10 to 40 credits depending on output size, a rate it marks as limited-time. Check your plan before you promise a quick turnaround.

Content Credentials attach attribution and edit history to an exported image, a kind of "nutrition label" in Adobe's words. Adobe adds them automatically to some exports that contain Firefly output, and says they show when a partner model was used. Its Generative AI User Guidelines (updated May 2026) are clear: you must not remove, alter or disable them.

Commercial use: what Adobe's own terms say

This is a summary, not legal advice.

• Commercial use. Adobe's Firefly FAQ (May 2026) says outputs from features without the beta label can be used in commercial projects, and beta outputs too unless the product says otherwise.
• Training data. Adobe says its Firefly models are trained on licensed content, such as Adobe Stock, and public-domain content, and not on Creative Cloud subscribers' personal content.
• Partner models. Adobe says they are not developed by Adobe and that you are responsible for deciding whether one suits your project. It also says your content is never used to train partner or Adobe models.
• Third-party rights. The user guidelines prohibit prompts designed to produce copyrighted or trademarked content, and uploading reference images that contain someone else's copyrighted work.
• Indemnification. Adobe's legal FAQ for enterprise customers says it indemnifies Firefly output for customers who have bought the appropriate entitlement. Read your own agreement before you rely on it.

Who owns a generated image is a separate question, and the answer depends on the country. I go through it in are AI images copyrighted?

From my work: where I draw the line

On packaging, the line is clear. The seven packs I designed for Twins Pharma each have their own colour and a clear bilingual label, and patients and pharmacists must be able to trust every word on it, so nothing generated belongs in that artwork. At most, a generated background could sit behind a pack in a presentation scene, labelled as such.

In the AI video series I direct at SkillUp MENA (AI Video Studio), image tools are one step in a chain of script, stills, motion, voice and edit. The same discipline applies there as in Photoshop: the tool proposes, and every frame is checked before it moves to the next step.

Conclusion

Generative Fill and Generative Expand are real time-savers for extending backgrounds, removing distractions, building scenes and retouching. Use them on their own layers, in small areas, with the brand's colours corrected by hand and every result checked at actual pixels. Keep text, Arabic letters, logos, labels and product shapes real, leave Content Credentials in place, and read the terms of the model you pick.

If you need brand, packaging or campaign visuals where AI speeds up the work and a designer stays responsible for every pixel, start a project here.