Adobe Put Generative Editing Inside Lightroom, and Photographers Are Right to Be Nervous

The September update to Lightroom desktop, version 9.6, added an early-access feature called Prompt to Edit. You describe the change you want in plain language, or pick one of the example prompts Adobe supplies, and the software applies it. The examples cover restoring colours, cleaning up dust and scratches, and reducing blur. The result is saved as a DNG next to the original file, and the interface tells you how many AI credits the operation consumed.

Adobe frames this as generative editing arriving inside Lightroom, so you no longer have to bounce a photo out to Firefly and back. That framing is accurate as far as it goes. What it leaves out is what the feature actually does to the file.

It regenerates the photo, not the region

The prompts suggest a targeted fix. The behaviour is closer to a repaint.

Thomas Fitzgerald, a photographer who writes about editing software, tried it on an image and asked for two things: brighten the frame, and add sunset colour to the sky. It did both. It also changed the railings in the shot, and added people who were not there when the photo was taken. What came back was a different photograph.

That is not a bug in the usual sense. The tool runs a generative model over the image, and a generative model has no obligation to keep the parts you did not mention. Adobe lists Nano Banana 2 output options at 2K and 4K for this feature. Nothing in the interface promises that the pixels outside your description will survive untouched.

The comparison worth making is with FLUX 3 Image, which Black Forest Labs released on October 1 and which ships with bounding-box controls. You place an element on a normalised grid, and the model only regenerates inside that box, leaving the rest bit-identical. Adobe's approach gives you a sentence and takes the whole frame.

The file format hides the cost

Prompt to Edit writes its output as a DNG, which sounds like it preserves raw flexibility. It does not. The saved DNG holds flat pixel information rather than sensor data, so the raw latitude your camera captured is gone once the operation runs. You get a file that looks like a raw container and behaves like a finished render.

There is a legitimate use for that. If a client wants a sky replacement on a stock image and nobody cares how it was made, the workflow is faster than masking. The problem is the framing. A tool that sits inside a raw editor, in the panel where photographers make routine adjustments, encourages the assumption that it is a routine adjustment. It is a generative rewrite with an undo step.

The credit meter changes the incentive

Adobe shows the credit cost before you commit. That detail matters more than it looks. Generative operations are billed, and a photographer working through a shoot of 400 frames has to decide, frame by frame, whether a generative fix is worth the credits. The pricing structure pushes users toward batching and toward accepting the first result.

It also pushes them toward treating the operation as cheap. When an edit costs a visible amount of money, the temptation is to justify it by using it. The tool is positioned to be used often.

Lightroom's earlier AI features behaved differently

Adobe has been shipping AI in Lightroom for several years, and those features kept the file intact.

Denoise AI removes noise and writes a new DNG that sits beside the original, with the adjustment non-destructive and reversible. Masking selects a subject and lets you change exposure on it without touching a pixel outside the mask. Super Resolution enlarges an image and keeps the result editable. All three change the image while preserving the relationship between the original capture and the edit.

Prompt to Edit breaks that pattern. The output is a new image, and the connection to the capture is interpretive rather than arithmetic. You cannot reverse it by moving a slider.

That is why the placement of the feature matters as much as the feature itself. Lightroom's panel layout teaches a workflow: adjust, mask, retouch, export. A generative operation sitting inside that sequence inherits an expectation of reversibility it does not deliver.

The same capability landed in Photoshop first

Adobe has already put generative editing into Photoshop, with the same caveats. There is a separate feature in the Photoshop beta called Chat to Edit, which works differently: instead of generating pixels, it drives the application's own interface through an agent. Fitzgerald tried that one and was unimpressed, calling it poor. The distinction is useful. Chat to Edit controls the tools. Prompt to Edit replaces the image.

Neither replaces the underlying judgement about whether a photo should be altered at all.

Where this leaves the working photographer

The honest split is between work that has to depict a specific moment and work that has to look right.

For news, documentary, evidence and any client assignment where the photograph is the record, a tool that silently adds people to the frame is not a tool you can use. You cannot easily tell, after the fact, which pixels the model invented. The saved DNG does not annotate the generated region. A colleague opening the file later sees a photograph, not a composite.

For commercial work where the image is a fabrication from the start, the calculus is different. A campaign shot built in a studio with a replaced background was already synthetic in places. Extending that to a replaced sky is a step, but it is a step along a path the industry already chose.

The gap is in the middle, where most freelance work sits. A portrait retouched with a generative pass that also moved a stray hand out of the frame is a better picture and a worse document. Photographers will make that trade individually, and the software makes it easier to make it without noticing.

What to watch

Three things will decide how much this spreads.

The first is whether Adobe adds provenance marking to generative output. The saved DNG carries no visible or embedded signal that a model touched it. Content Credentials would change that, and Adobe sits on the standards body that defines them.

The second is whether the interface separates generative edits from adjustments more clearly. Right now Prompt to Edit lives in the same space as sliders that change exposure by a third of a stop. Those two operations do not deserve equal billing.

The third is what other editors do. ON1 shipped NoNoise AI 2027 on October 1 with rebuilt denoising models. DxO PhotoLab 10.1 added direct export to Affinity's native format. Both are additive features that leave the original file intact. If the market rewards Adobe's approach, and if competitors copy it, the expectation that an editor preserves the photograph rather than replacing it will erode.

The feature is genuinely useful. The question is whether the people using it will know which photograph they are looking at when they are done.