Borrowed vs. Hallucinated Pixels

Professional visual work relies on non-destructive editing. This is the practice of making changes to an image without overwriting the original data. If you flatten your image or edit directly on the base layer, you lose the ability to revert a mistake or adjust a detail later. In a professional environment, your original source is your only absolute truth; it is the only part of the file you can guarantee is an accurate representation of the physical product.

For example, imagine you have a high-resolution studio photograph of a product bottle. In your Photoshop PSD, this photograph should be the very first layer at the bottom of the stack. Instead of painting or erasing on this layer, you lock it. By clicking the lock icon in the Layers panel, you create a safety barrier. Every subsequent change—whether it is a retouching fix or a generated background—happens on a new layer above it. This ensures that the raw, photographed pixels remain untouched, allowing you to ship a set of variants while always having a clean reference to verify against.

Not all automated tools create new content. Traditional retouching tools, such as the Clone Stamp and the Healing Brush in Photoshop, perform what is known as 'borrowing pixels.' These tools do not invent new imagery; instead, they sample existing data from one part of your photograph and move it to another. Because they rely entirely on the pixels already present in your source file, they maintain the original texture, lighting, and grain of the photo.

Consider a scenario where you are retouching a product label on that same studio bottle. You notice a tiny dust speck on the white background of the label. To remove it, you select the Clone Stamp tool and 'sample' a clean, dust-free area of the white label nearby. When you paint over the speck, Photoshop simply copies the clean pixels from your sample point to the target area.

As a professional, you use this distinction to maintain a 'paper trail' of authenticity. When you borrow pixels, you are rearranging the truth of the photograph. You aren't adding something that wasn't there; you are simply cleaning the frame using the image's own visual DNA.

While the clone stamp borrows data, Generative AI 'hallucinates' pixels. Hallucination occurs when a tool produces new visual content based on statistical patterns learned from millions of other images, rather than sampling from your specific source file. These pixels never existed in your photograph; they are an AI's best guess at what a specific object or environment should look like. This is a powerful shift from retouching to creation, but it introduces a new risk: the AI does not understand physics, geometry, or the specific product you are selling.

For example, if you take that studio bottle and use Generative Fill in Photoshop to place it on a stone table outdoors, the AI isn't looking for a stone table in your original file. It is inventing one. It generates the grey texture of the stone, the dappled sunlight of a summer afternoon, and the blurred greenery of the background from scratch.

As a professional, you must recognize that every 'hallucinated' pixel is a variable that needs verification. Unlike borrowed pixels, which are tied to the photo's reality, hallucinated pixels are predictions that can easily fail at a detailed level, such as where the bottle meets the table.

A dangerous trap in AI-assisted production is the 'convincing impression.' This happens when an image looks polished and professionally lit at a glance, leading you or your client to believe it is finished. Because generative AI is a statistical engine, it excels at creating a plausible overall atmosphere, but it often fails at the precise points where two different elements meet. The image may pass a brand-scoring tool or a quick visual review because the colors and mood are correct, but the craft is fundamentally broken in the details.

For example, look at your generated stone table background. At 100% zoom, the lighting looks natural and the texture of the stone is convincing. However, when you zoom in to 200%, you might see that the base of the bottle doesn't actually touch the stone. Instead, there is a blurred, nonsensical smudge of grey pixels—a 'hallucination'—where a sharp contact shadow should be.

As a professional, you must resist the urge to trust the overall impression. Your job is to hunt for these failures by inspecting the physics of the image at high magnification. You identify these errors not by how the image 'feels,' but by checking if the pixels obey the laws of light and geometry.

Once you have a mix of borrowed and hallucinated pixels, you can no longer rely on a simple filename to describe the file's contents. Professional workflows require a map—a clear record of which image areas are photographed and which are generated. Without this map, you risk losing track of what is a factual representation of the product and what is an AI guess. If a client asks for a change to a specific detail, you need to know instantly if you are editing a real photograph or a hallucination, as the tools and the level of caution required for each are different.

To do this, you create a layer manifest. This is essentially a directory of your Layers panel. For example, in your olive oil project, you would explicitly label your layers in the PSD: the base bottle layer is marked as 'Photographed,' while the added olive branch and the stone table are marked as 'Generated.'

You use this manifest as a guide during the review process. When you send a proof to a stakeholder, this mapping allows you to defend your craft decisions. You can point to the 'Photographed' areas as the absolute source of truth and the 'Generated' areas as creative additions that are subject to revision.

The internal map you create in your PSD is for your eyes, but the exports you send to clients need a different kind of transparency. When a file passes an automated brand check, a client may assume the entire image is a factual photograph. If that image is later used in a legal or regulatory context—such as a product ad—and the image contains hallucinated details that don't match the real-world item, the advertiser is held responsible. Clear labeling at the export stage prevents the dangerous assumption that a 'polished' image is a 'pure' one.

To manage this, you must implement a naming convention that distinguishes the source. For example, instead of naming a file OliveOil_Hero_v1.jpg, you use a suffix: OliveOil_Hero_v1_PurePhoto.jpg for images containing only photographed pixels, and OliveOil_Hero_v1_Hybrid.jpg for those containing generative AI. You can also add this status to the image metadata tags. By doing this, you create a permanent record that travels with the file, ensuring that any stakeholder who receives the asset knows exactly which parts are unreviewable AI guesses and which are the absolute truth of the product.

The final step in protecting your craft is establishing a hard boundary in the creative brief. Because AI tools are designed to be 'helpful,' they often overstep by altering details you never intended to change. If you simply ask for a 'natural outdoor setting,' the AI might subtly shift the gold of the olive oil or smudge the text on the label to fit the new lighting. To prevent this, you must explicitly categorize every element of the image as either 'Required' or 'Optional' before you ever open a generative tool.

For example, in your olive oil project, you define the product cap, the label text, and the specific color of the oil as 'Required/Untouchable.' These are the factual anchors of the product that must match the bottle on the shelf. In contrast, the ambient lighting, the stone table texture, and the decorative olive branch are marked as 'Optional/Generated.'

By documenting this boundary, you transform a vague visual 'feeling' into a reviewable contract. If a client's AI tool approves a version where the label is slightly blurred, you can point to the brief to justify your pushback: the label was marked 'Untouchable,' and therefore the file is a failure regardless of the AI score.