Reference Photos for AI Scenes: Light, Background, Angle

You upload a product photo, ask for a cozy kitchen scene, and the mug comes back a shade too orange with a handle that wasn't there before. The instinct is to regenerate and hope. Usually the better move is to look at the photo you fed in.

This article covers what makes a good reference photo, how to shoot one with a phone and a window, and what to change when a scene still comes out wrong.

Why the reference photo decides the scene

An image tool that works from a reference picture treats that picture as the truth about your product: its shape, colour, material, logo and proportions. Google's developer documentation on image generation describes using an input image to guide what the model produces. The input steers the output. Anything the photo leaves unclear, such as a hidden side, a blurry logo or a color shifted by bad light, the tool has to guess.

It helps to separate two jobs:

The reference photo describes the product. It should be accurate, clear and boring in the best sense.
The scene is built around the product. Mood, setting and props come from your scene description and style, not from the reference.

So a beautiful photo can still be a poor reference. A moody shot with deep shadows, a warm filter and the product half in the dark looks great on a feed, but it tells the AI very little about what the product really looks like. A flat, evenly lit shot on a plain surface looks dull and works far better.

What follows is a short set of habits for light, background and angle, then a repeatable five-minute routine and a problem-to-fix table.

Light: even, soft and honest about colour

Light is the biggest factor you control, and it costs nothing.

Use window light or a shaded spot. Put the product beside a window, out of direct sun, and shoot from the side where the light falls. Turn off the phone flash, which creates harsh highlights and flat, unnatural color. Also switch off the room lamps. Mixed light sources (a window plus a warm bulb, for example) push colors in different directions on the same product.

Soften hard shadows and glare. Glossy, metallic and glazed products reflect whatever is around them. A hard highlight or a dark reflection can be read by the AI as part of the object, and then it shows up in every scene. The simple fix is a white card or a plain sheet of paper placed on the shadow side of the product. It bounces light back, fills the shadows and softens the highlights. Move it closer or farther until the shadow looks gentle.

Protect the colour. Customers return products that don't match the picture, so color accuracy matters twice: once in the reference, and again in the final post. After shooting, hold the real item next to your screen and compare. If the screen version is warmer, cooler or more saturated than the object, reshoot or correct it before you upload.

Keep editing minimal. Heavy filters, strong contrast boosts and HDR looks bake in distortions the AI will treat as facts. If you edit at all, limit yourself to straightening, a light exposure fix and a small white-balance correction.

Tip

A quick test for light: if you can't tell where the shadow starts and ends, it's soft enough. If you can draw its outline with a pencil, add the white card.

Background: plain, contrasting and clean at the edges

The tool has to work out where your product ends and the rest of the picture begins. Make that easy.

Choose a plain, uncluttered surface. A sheet of paper, a table, a piece of fabric without a busy pattern. Then pick a color that contrasts with the product. A white ceramic mug on a white sheet gives the tool almost no edge to follow; the same mug on a mid-gray or soft colored backdrop has a clear outline.

Keep the edges clean.

No props touching or overlapping the product.
No cut-off corners at the edge of the frame.
No hands, fingers or reflections that blend into the outline.
No shadows from the product merging into a dark background.

Give tricky materials extra help. Transparent items (glass, clear plastic), fine hair, fringe, lace and mesh are hard for any tool, because the background shows through them. A contrasting backdrop helps most here. For a glass bottle, a dark backdrop with a light side card shows the shape; for a dark mesh bag, use a light backdrop.

Remove anything you don't want copied. Price tags, stickers, old watermarks, store badges and sale banners can be reproduced in the new scene. This also matches what marketplaces expect: Google's product image guidance for Merchant Center asks for clear product images without overlays or watermarks (as of October 2026, check the page for the current detailed requirements). Shopify's help on product media also makes the point that images are shared across your store, so a clean, consistent set serves both the product page and your AI scenes.

Angle and framing: show what a buyer would need to see

Shoot from the angle a customer would naturally look at the product. For a mug, that's slightly above and to the side; for shoes, from the side at about shoe level; for a poster or a tray, straight on. Keep the camera level and don't tilt it. A tilted phone stretches proportions: a tall bottle looks squat, a round plate looks oval.

Fill the frame sensibly. The whole product should be visible, with some air around it. If a corner or the base is cropped, the AI will invent the missing part, and the invention rarely matches your real item. Standing a little farther back and cropping later is safer than getting too close.

Take a small set, not one picture. One angle can't describe a three-dimensional object. A useful minimum:

Front view.
Side view (or back, if it differs).
A close-up of the key detail, such as a logo, stitching, a clasp or a texture.
One shot that shows scale: held in a hand, worn, or next to a familiar object.

Keep text and logos flat and sharp. Image tools often distort lettering. A logo photographed at an angle, or slightly out of focus, comes back warped. Shoot labels straight on, in sharp focus, and include one dedicated close-up. If a scene still bends the lettering, you may be better off keeping the product photo as is and placing the text on a real label or post caption instead.

A worked example: the ceramic mug

Take a speckled white mug with a small printed logo on one side.

First attempt: shot at night under a ceiling lamp, on a white tablecloth, handle turned away, logo hidden, phone flash on. The result is a mug that looks yellow and has a harsh glare stripe. The AI scene puts the glare into the mug as a "shiny band", gives it a different handle and makes up a logo.

Second attempt: shot beside a window on a mid-gray cloth, white card on the shadow side, handle pointing right, no flash. Three angles plus a straight-on logo close-up. The color matches the real mug on screen, the handle is clear and the logo is readable. The scene comes back with the right mug on the first or second try, and the time you saved far exceeds the five minutes spent shooting.

A five-minute reference shoot you can repeat

Here is a routine to run for each product. Once you have a spot by the window, it really does take about five minutes.

Step by step

Clean the product. Wipe off dust, fingerprints and smudges. Remove tags and stickers.
Set up. Place the product on a plain, contrasting surface beside a window, out of direct sun. Add a white card on the shadow side.
Shoot the set. Front, side, detail close-up, and one scale shot. Keep the phone level and the product fully in frame.
Check. Zoom in on each photo: is the edge sharp, the logo readable, the color right against the real item?
Name and upload. Use clear file names such as mug-blue-front.jpg, mug-blue-logo-closeup.jpg. Add them to your media library; SemPost's help page on uploading pictures and videos explains how.

Pass/fail checklist (copy it into a notes app):

If any line fails, fix it and reshoot that one photo; don't push on and hope the tool compensates.

Note

If you also sell through product catalogs or marketplaces, check the image requirements on each platform's own help pages before you reuse the same files. Rules differ by channel and change over time, and the Merchant Center guidance is a good starting point for clean listing images.

When the files are in your library, you can pick them as the source for scenes. The help page on generating posts from your products walks through that step.

When the scene looks wrong: a problem-to-fix table

Even with a good reference, some results will miss. The trick is to work out where the problem started. Use this table to match what you see to its likely cause and a fix.

What you see in the result	Likely cause in the reference	What to change
Wrong colour (too warm, too cold, washed out)	Mixed or warm light, flash, a filter or heavy editing	Reshoot in window light without filters; compare to the real item on screen
Warped or invented logo or text	Logo at an angle, blurry, or small in the frame	Add a sharp, straight-on close-up; keep the logo flat to the camera
Extra parts (second handle, added straps, invented buttons)	A hidden or cropped side, unclear shape	Add a side or back angle; keep the full product in frame
Product floating or wrongly scaled	No sense of size in the photo; a tilted camera	Add a scale shot (in hand, next to a familiar object); shoot level
Edges melting into the background	Low contrast with the backdrop; reflections or props touching	Reshoot on a contrasting, plain surface with clean edges
Glare or a shiny band that wasn't on the product	Hard light or reflections in the reference	Soften the light with a white card; move away from reflective surfaces
Old price tag or watermark appears	It was visible in the reference	Remove it from the product or choose a clean image
Right product, wrong mood or setting	The scene description or style, not the photo	Adjust the scene prompt or style, not the reference

Reshoot or re-describe? If the product itself is wrong (color, shape, logo, extra or missing parts), the fix lives in the reference photo. If the product is right but the setting, mood, props or lighting of the scene are off, change the scene description or style instead. The help pages on scene styles, tones and languages and on rewriting the text or making a new image cover those controls, and the generation troubleshooting page lists common problems.

A short example of a scene description kept focused on the setting only:

Place this exact mug on a wooden kitchen counter next to a window,
morning light, a folded linen towel in the background.
Keep the mug's color, handle and logo exactly as in the photo.

Change one thing at a time. If you swap the photo, the style and the description all at once, you won't know what helped. Change the reference first if the product is wrong, then rerun with the same description. Change the description only when the product is right.

Watch out

Before you post, compare the final picture with the real product. A scene that looks great but shows the wrong shade, an extra feature or a distorted logo can lead to disappointed customers and returns. Also check whether the picture needs an AI label: some networks have their own rules for AI-generated images, and SemPost's page on AI labels explains how labeling works there. This is general information, not legal advice; if you are unsure about disclosure rules where you sell, check the platform's own help pages or ask a qualified professional.

What to do today

Pick your five best-selling products and reshoot just one thing for each: the light. Move them next to a window, add a white card and drop the flash. Then shoot the three extra angles for the product with the most logo or detail. Run the checklist, upload the files and generate a scene from each. Compare them with your older results; if they match the real products better, the habit has paid for itself.

A reference photo takes minutes to improve and is reused every time you make a scene from that product. That is why it deserves your attention before the next regenerate click.