Borrowing vs. Inventing Pixels

Professional retouching relies on non-destructive editing: a workflow where no original pixel is ever overwritten. In Adobe Photoshop, the foundation of this habit is the locked background layer. When you open a source file, your first action should be to lock the original image at the bottom of the layer stack. This ensures that every change you make exists on a separate layer above the source, allowing you to revert any decision without restarting the project.

Consider a studio shot of an olive-oil bottle on a grey seamless background. If you begin painting or filling directly onto the base image, you destroy the original data. Instead, you lock that background layer immediately. If you later decide to change the stone tabletop or the lighting, those additions sit in their own layers. If a client asks to return to the original studio lighting, you simply hide the generative layers to reveal the pristine source. By keeping the original locked, you separate the truth of the photograph from the iterations of the edit, ensuring the work remains reviewable and reversible.

Not every tool in Photoshop creates something new. Traditional retouching tools, such as the Clone Stamp and the Healing Brush, operate by borrowing pixels. This means they take existing texture and color from one part of your image and move it to another. These tools do not 'think' or imagine content; they simply replicate what is already there. Because they rely on existing source data, the result is a precise copy of the original photograph's grain, lighting, and detail.

Imagine you are cleaning up a studio shot of a stone tabletop. You spot a small, distracting dust speck on the surface. To remove it, you select the Clone Stamp tool and Alt-click a clean, texture-matched patch of stone just an inch away from the speck. When you paint over the dust, Photoshop copies the pixels from that clean patch and pastes them over the speck. You haven't invented new stone; you have simply repositioned existing stone to hide a flaw. When you use these tools, your goal is to ensure the borrowed texture blends seamlessly with the surroundings without creating an obvious repeating pattern.

While the Clone Stamp borrows pixels, Generative AI creates them. This process is often called 'hallucinating' pixels because the tool produces content that was never present in the original photograph. Instead of copying a nearby patch of texture, Adobe Firefly uses a statistical engine to guess what a specific object should look like based on millions of other images it has seen. It does not understand the physical laws of the scene; it only knows that certain patterns of pixels typically represent certain objects.

For example, if you want to add a fresh olive branch to your stone tabletop, you wouldn't use the Clone Stamp unless a branch already existed elsewhere in the frame. Instead, you select an area and use Generative Fill. The AI 'invents' the organic curves of the stems and the specific veins of the leaves. These pixels are synthetic; they are a mathematical best-guess of an olive branch. As a retoucher, your job is to recognize that these pixels are invented. Because they are hallucinations, they are prone to physical errors—like a leaf merging into the stone or a stem that defies gravity—which requires a rigorous inspection routine at 200% zoom.

Once you begin mixing borrowed and invented pixels, the image becomes a hybrid of reality and statistical guesses. If you don't track which is which, you risk treating a hallucinated detail as a factual reference. Professional retouchers use a layer manifest—a clear labeling system in the Layers panel—to map these sources. This audit trail ensures that anyone reviewing the file knows exactly which pixels are anchored in the original photograph and which were generated by Firefly.

Take our olive-oil bottle project. You start with the studio shot of the bottle, which you label as 'Product_Photographed'. You then use Generative Fill to create a rustic stone table and a summer afternoon background, labeling these layers as 'Setting_Generated'. If you add an olive branch, that becomes 'Detail_Generated'.

By explicitly naming your layers this way, you create a map of the 'hallucinations' in your scene. When it comes time for the final inspection, you don't waste time checking if the bottle's label is accurate—you know it's photographed. Instead, you focus your 200% zoom inspection on the 'Generated' layers, where the AI is most likely to have failed the laws of physics.

In Photoshop, 'flattening' happens when you merge multiple layers into one single background layer. While this reduces file size, it is a critical error in a generative workflow. When you merge a generative layer into your base image, you turn a machine's guess into a permanent part of the original source. Once those pixels are flattened, they are no longer separate from the photograph; they become unreviewable and nearly impossible to adjust without destructive cloning or masking.

Imagine you use Generative Fill to add a reflection of your olive-oil bottle onto the stone table. The AI produces a convincing image, but you immediately merge that reflection layer into the base photograph to 'clean up' your Layers panel. Later, the client zooms in and notes that the reflection's angle is slightly off—it doesn't align with the bottle's perspective. Because you flattened the layer, you cannot simply transform or warp the reflection. You are forced to destroy parts of your original stone texture to fix a machine's mistake.

To avoid this, never merge generative content into your source. Keep every 'hallucinated' element on its own named layer until the final export. This ensures that if a generated detail fails a visual review, you can delete or adjust it without touching a single original pixel.

During a technical review, your eye must be able to distinguish between a borrowed pixel and an invented one. This is not about whether the image looks 'good,' but about identifying the logic used to create the pixels. Borrowed pixels are based on duplication; invented pixels are based on synthesis. This distinction tells you exactly what kind of error to look for when you zoom in to 200%.

Consider two areas of a stone tabletop. In the first area, you used the Clone Stamp to cover a scratch. If you look closely, you will see a repeating pattern—the same unique pit or grain of stone appearing twice in the same area. This is a 'borrowed' error. In the second area, you used Generative Fill to extend the table. Here, there are no repetitions, but you might find 'synthetic' errors: a stone edge that blends unnaturally into the background or a texture that looks like smoothed plastic rather than mineral. By identifying the source, you know whether to fix a repetition with the Healing Brush or to re-generate a hallucination that defies physics.

The final safeguard against generative errors is the audit trail. This is the systematic organization of your PSD file so that every generative task is isolated. Instead of using one large generative area to build a scene, you treat each addition as a discrete task. This creates a clear path back to the original source, allowing you to toggle individual elements on and off to see how they interact with the original photograph.

In our olive-oil bottle project, you don't simply 'generate a background.' Instead, you build the scene in targeted layers. First, you create a layer named 'Stone Table Gen' for the surface. Above that, you add 'Evening Light Gen' to handle the warm atmospheric glow. Finally, you add 'Olive Branch Gen' for the decorative detail. All of these sit above your locked original studio shot.

By organizing your file this way, you transform your Layers panel into a checklist. If the client decides the olive branch is too distracting, you delete one layer without affecting the lighting or the table. This structure ensures that you never lose the original 'truth' of the image while maintaining total control over every hallucinated detail.