Photoshop Generative Fill: What It Still Cannot Do

10:14

Most articles about Generative Fill alternatives are out of date, and in a specific way that matters: they assume Generative Fill can only use Adobe's own Firefly model. That stopped being true.

Photoshop now offers partner models inside the Generative Fill workflow — you pick from a model selector in the Contextual Task Bar, grouped as Adobe or Partner. The partner list currently includes Gemini 3.1 (Nano Banana 2), Gemini 3 (Nano Banana Pro), Gemini 2.5 (Nano Banana) and FLUX.2 pro, with FLUX Kontext Pro also available.

That change quietly resolved the most common complaint about Generative Fill, which was never really "the feature is bad" — it was "the model behind it is not the one I want." So this guide covers what is genuinely fixed, the four things Generative Fill still cannot do, and how to decide which side of the line a job belongs on.

What is Photoshop Generative Fill good at?

Selection-based work: removing an object, extending a background, filling an awkward gap, or replacing a small region while the rest of the file stays untouched. It operates on your actual layered document at your actual resolution, with no export and no round trip. Nothing outside Photoshop matches that.

It is worth being precise about the strengths, because the round-trip alternatives genuinely lose all of them.

• It respects your selection. You define exactly which pixels may change. A standalone model edits the whole image and you hope the rest survives.
• It stays in your document. Layers, masks, smart objects and colour profile are all intact. No export, no re-import, no mismatch.
• It is genuinely good at small removals. A sign, a tourist, a power line, a blemish on a wall. For this category it is fast and usually right first time.
• Background extension is excellent. Widening a crop or filling a corner after a straighten is one of the cleanest things it does.
• Adobe's commercial terms. Adobe states your work stays yours and is not used to train AI, including when you choose a partner model. For agencies with procurement requirements, that matters.

Which AI models can Generative Fill use now?

Adobe's own Firefly, plus partner models selected from the Contextual Task Bar. The partner list currently includes Gemini 3.1 with Nano Banana 2, Gemini 3 with Nano Banana Pro, Gemini 2.5 with Nano Banana, and FLUX.2 pro, with FLUX Kontext Pro also offered. The selector groups them as Adobe or Partner.

This is the single most important update for anyone who dismissed Generative Fill a year ago. The models people were leaving Photoshop to reach are now, in large part, available inside it.

Partner model availability verified against Adobe's documentation in August 2026. Adobe adds and rotates partner models, so check the selector in your version before assuming a model is missing.

The gap that remains is a curated-list gap, not a quality gap. If the job needs accurate in-image text, GPT Image 2 is materially better than anything in the selector — and it is not in the selector.

Why does Generative Fill still fail on some jobs?

Four structural reasons: it fills a selection rather than editing a whole image, it works one file at a time, the generated region is rendered at a capped resolution rather than your document's, and the model list is curated. None of these are quality problems — they are shape-of-the-tool problems.

1. It fills selections, not intentions

"Remove this sign" is a selection. "Relight this portrait as if the key light came from the left, keep her face identical" is not — it is an instruction about the whole image. Generative Fill has no way to express that, because its unit of work is a region you drew.

2. One image at a time

Photoshop is a single-document tool. Applying the same edit to sixty photographs means sixty rounds of select, prompt, generate, accept. Standalone models take a batch and work through it while you do something else.

3. Resolution of the filled region

The generated patch is produced at a bounded resolution and scaled into your selection. Adobe has raised that ceiling over time and it is far less noticeable than it was, but on a large selection in a high-resolution file the filled area can still read softer than the surrounding pixels. Working outside, you generate natively at 2K or 4K and composite at full quality.

4. The model list is curated

Excellent as the partner list is, it is a list. GPT Image 2 for text, Seedream for imaginative work, and the cheap fast models for exploration are not on it.

What does Generative Fill cost?

It draws on generative credits included with your Creative Cloud plan. The allowance and the overage terms vary by plan and Adobe has revised them more than once, so check your plan rather than trusting any article's number. Pay-per-image models outside Photoshop run two to nine cents with no monthly floor.

I am deliberately not quoting an Adobe credit figure here. The allowances differ by plan, differ by region, and have changed — any specific number in a blog post is likely to be wrong by the time you read it. Your Creative Cloud account page is the only reliable source for yours.

What can be stated concretely is the other side of the comparison, because those rates are published per image:

Rates via Kie.ai at 2K, August 2026, checked against Rangy's live pricing tables. Rates vary by provider and change over time.

When should you work outside Photoshop?

When the edit applies to the whole image rather than a selection, when you have a batch, when you need native 4K output, or when the right model is not in Adobe's partner list. Those four cases are where a round trip earns its friction — and nothing else is.

• Whole-image instructions. Relighting, style transfer, "keep her face identical but change everything behind her", restoration of a badly damaged photograph. These are prompts about an image, not a region.
• Batches. Sixty product shots that all need the same background treatment. One instruction, queued, versus sixty manual rounds.
• Native high resolution. When the edited area is large and the file is going to print, generating at 2K or 4K natively and compositing beats scaling a filled patch.
• Models outside the list. Readable in-image text is the clearest case — GPT Image 2 is meaningfully better at it and Adobe does not offer it.

How do you move an image out and back?

Make your selection in Photoshop and copy it, paste it into the external app, generate with the model you want, then paste the result back onto its own layer. Copying a selection rather than the whole document means only the region you care about makes the trip, and it returns aligned.

The friction is real, and it is the honest price of stepping outside. Four steps and two application switches against one Contextual Task Bar prompt. That is why the four cases above are the only ones worth doing it for.

A detail that saves rework: paste the returned image onto a new layer rather than over your original, then mask it in. If the model shifted something you did not want shifted, you still have the original underneath.

Which model should you use for which fix?

Stay in Generative Fill for object removal and background extension. Use Nano Banana Pro when a face must stay identical, GPT Image 2 when the image contains readable text, and a cheap fast model when you are exploring rather than finishing.

The Nano Banana 2 vs Pro comparison covers the editing pair in detail, and the portrait retouching guide has the prompt wording that keeps either model conservative on faces.

Is there a free alternative to Generative Fill?

Locally-run open-weight models with an inpainting workflow are genuinely free after hardware. Free browser tiers exist but watermark or cap output. If you already subscribe to Creative Cloud, the honest answer is that Generative Fill is the cheapest option you have for the jobs it handles.

Local open-weight models such as Stable Diffusion or FLUX with an inpainting interface cost only electricity and keep everything on your machine. The price is a capable GPU and setup patience.

Free browser tiers are useful for evaluating a model before committing, but generally watermark output or cap resolution. The free AI image generator guide covers which permit commercial use.

Already paying for Creative Cloud? Then the cheapest thing you can do is use the credits you have already bought. Reaching for a paid alternative to do a job Generative Fill does well is spending money twice.

When should you just use Generative Fill?

For any selection-based fix on a single file: removing objects, extending backgrounds, filling gaps after a straighten or crop. It is faster, it keeps your layers intact, it is already paid for, and since the partner-model update its output quality is competitive with what you would get outside.

I sell a tool that competes for some of this work, so this section is the one to read most sceptically — but four cases genuinely belong inside Photoshop.

• Small removals. A round trip to remove a lamp post is pure overhead. Select, prompt, done.
• Anything where layers matter. If the edit needs to sit in a specific place in a layer stack, flattening and re-importing costs you more than the model gains.
• You have credits sitting unused. You already paid. Use them before paying per image elsewhere.
• Client work with procurement requirements. Adobe's stated terms on ownership and training are a real procurement advantage, and reproducing that assurance across several third-party providers is harder.

The realistic setup for most people is both: Generative Fill for selection work inside the document, and an external app for whole-image edits, batches and the models Adobe does not carry. They are complementary rather than competing.

Frequently asked questions

The bottom line

Generative Fill got much better when Adobe opened it to partner models, and most articles telling you to replace it have not caught up. Use it for everything selection-shaped. Step outside only for whole-image edits, batches, native high resolution, or a model Adobe does not carry.

The old framing — Adobe's model versus everyone else's — is finished. Nano Banana Pro is available in both places now. What remains different is the shape of the tool: a selection filler inside your document, versus a whole-image generator that batches.

Pick by the shape of the job, not by the model name. Most weeks, that means staying in Photoshop more often than an article like this one would normally admit.