How to Use Reference Images with Generative Fill in Photoshop

How to Use Reference Images with Generative Fill in Photoshop

Describing what you want in words only gets you so far. Ask Generative Fill for sunglasses and you will get sunglasses — just not the ones you had in mind, and the odds of arriving at a specific style through prompt wording alone are poor.

Reference images solve that. Instead of describing the thing, you show Photoshop a picture of it, and the generation takes its design cues from what you supplied. It is a substantial change in how much control you have, and it comes with one significant catch worth understanding before you start.

A note on versions: this was recorded when reference images arrived as a beta feature in the Photoshop beta build. It may now be in the main release, and the interface may differ from what is described here. The behavior and the workaround below still apply.

The Workflow

1. Make a selection with the Lasso Tool

   Draw around the area where the new object should go. Hold Alt or Option to subtract anything you caught by mistake.

2. Click Generative Fill

   From the contextual task bar.

3. Type a text prompt describing what you want

   Keep it short and concrete — “sunglasses”, “necklace”.

4. Click the reference image button and upload your reference

   A thumbnail appears alongside the prompt field.

5. Click Generate

   Variations come back, taking their design from the reference rather than from the model’s general idea of the object.

You Still Need the Text Prompt

The reference image supplements the prompt rather than replacing it. Uploading a photo of sunglasses without typing anything does not reliably produce sunglasses — the text tells Photoshop what kind of thing to generate, and the reference tells it what that thing should look like.

Think of the two as answering different questions. The prompt names the object; the reference supplies the style, shape and material. Both matter, and dropping either one costs you control.

The Catch: It Uses Everything It Can See

This is the thing to know before you waste a dozen generations working it out.

Photoshop reads the whole reference image, not just the object you were interested in. Upload a photo of a person wearing sunglasses and a necklace, and the generation will often pull design elements from both — you get sunglasses with jewelry-like detailing, or a hybrid of the two objects that is not what you asked for.

It has no way of knowing which part of your reference you meant. It takes the visual character of the image as a whole.

The fix: isolate the object first

Prepare the reference before you use it. Open your source photo, cut out just the object you want — sunglasses on their own, the necklace on its own — and save that as a separate file with everything else removed.

Feed Photoshop that isolated version and the results improve dramatically, because there is nothing else in the frame for it to borrow from. This one step of preparation is the difference between the feature being frustrating and being genuinely useful.

A clean cutout on a plain background is ideal. It does not need to be a perfect mask — the AI is reading design characteristics, not compositing your reference directly — but it does need to contain only the thing you want.

Building from more than one reference

The same behavior that causes the problem can be used deliberately. If you genuinely want to combine characteristics — the frame shape from one pair of glasses and the lens tint from another — a reference containing both will blend them.

That is a less predictable way to work, but for concepting it can produce ideas you would not have arrived at by describing them.

Swapping the Reference Without Starting Over

Once a generation exists, you can replace the reference image and generate again without redrawing the selection. There is also an option to remove the reference entirely and fall back to text alone.

This makes comparison quick. Keep the same selection and the same prompt, cycle through three or four different reference images, and you can see how each one influences the result — which is by far the fastest way to develop a feel for what the feature actually does with a reference.

Where This Is Genuinely Useful

Matching a specific product. If a client sells a particular frame, a particular watch, a particular garment, describing it in words is hopeless. A reference gets you close.

Consistency across a set. Using the same reference across several images keeps the generated element broadly consistent between them, which text prompts alone will not do.

Style direction. A reference conveys material, era and finish — brushed metal, worn leather, high-gloss plastic — far more efficiently than a prompt describing them.

More Free Photo Editing Tutorials

If this was useful, these go further in the same direction:

• How to Create a Composite in Photoshop — building the layered file first.
• Compositing Highlights, Shadows and Reflections — making an added element sit in the scene.
• How to Use the Select and Mask Workspace — cleaner cut-outs, which compositing depends on.
• How to Create and Use Layer Masks — confining an effect to exactly where it belongs.
• How to Make Realistic Shadows in Photoshop — the shadow that anchors an object.
• How to Use Select Subject in Photoshop — the selection engine behind most cut-outs.

Limitations Worth Knowing

It is influence, not insertion. Photoshop is not compositing your reference into the image. It generates something new that resembles it. Expect a family resemblance rather than an exact copy, which means this is not a way to place a real product accurately into a photo.

Resolution still applies. Generated content comes back at a capped resolution, so keep selections tight and check the result at 100% against the surrounding photograph.

It works best on distinct objects. References containing a clear, well-lit single object outperform busy scenes, subtle textures, or anything where the interesting part is small within the frame.

The Practical Habit

Build a small folder of isolated reference cutouts for the things you generate often. Ten minutes spent cutting out a handful of objects on plain backgrounds pays for itself the first time you need a specific style, and it turns a feature that is unpredictable out of the box into one you can direct.

That is the underlying point here. Generative tools reward preparation far more than they reward prompt wording — a clean reference will beat a cleverly phrased description every time.