AI Product Photography: Which Models Keep Your Product Accurate?

We gave seven AI image models the same phone-style photo of an amber bottle. Each made a packshot, a café scene, and a hand shot. All 21 kept the three label lines, but none produced a packshot with a fully pure white background. Nano Banana 2 came closest on white and darkened the amber glass nearly black; GPT Image 2.5 kept the color closer but framed the cap too tightly. Each edit cost from $0.0324 to $0.1226 in our run.

The comparisons below show the actual images, billing, and checks you need before using an AI product photo in a listing. We ran every edit through Modellix with one API key.

Which AI model should you use for product photography?

What matters most	Best pick in our test	Billed per image	Check before publishing
Whitest packshot background	Nano Banana 2 Edit	$0.0612	Amber glass turned almost black
Large framing with close amber color	GPT Image 2.5 Sunburst Edit	$0.0432	Cap sat 15 px from the top edge
Lowest edit cost	Qwen Image 3.0 Edit	$0.0324	Background averaged RGB 250
Fastest edit	Nano Banana 2 Edit	$0.0612	About 11 s per image
Square photo to 16:9 banner	Kling Image Expansion	$0.0224	Output measured 1360 × 768

Make product images you can actually review

Promotional composite using three images from this test. The full seven-model comparisons follow.

• Compare models on your own product. Run the same source photo through image models with one Modellix API key, then choose by label, color, and crop.
• Build a repeatable catalog workflow. Generate packshots and lifestyle variants from a SKU list, with the same review checks for every item.
• Know the cost of each image. Every completed call reports its billed amount; our seven-model edits cost from $0.0324 to $0.1226 each.
• Carry the approved shot into video. Use the finished product image as the starting frame for an image-to-video model on the same platform.

Try your product photo on Modellix

How did seven AI models edit the same product photo?

The source image and test setup

The source photo is AI-generated. It shows a fictional 500 ml HALDEN Cold Brew Concentrate bottle in warm kitchen light, with glass glare and a busy counter. We chose amber glass and a three-line label to make color and text changes easy to spot.

Every model got the same three prompts, one call each, no retries, at 1K resolution and 1:1 where the model has an aspect ratio setting. Each prompt ended with the same instruction:

Keep the bottle shape, the black cap, the label design and every word on the label exactly as in the original photo. Do not add any other text or logos.

We read every label back with on-device OCR (Apple Vision) and measured the packshot backgrounds pixel by pixel.

Which model made the best packshot?

Turn this phone photo into a clean e-commerce product photo: the same bottle alone on a pure white background, soft even studio lighting, a gentle contact shadow under the bottle, straight front view, centered.

All seven removed the kitchen. What they did with the bottle and the white was less even. Nano Banana 2 Edit gave the cleanest white but rendered the glass almost black, which would misrepresent an amber bottle on a product page. GPT Image 2.5 Sunburst Edit made the bottle fill the frame, close to what a marketplace main image wants, but cropped it tight at the cap. Wan 2.7 Image Pro Edit returned a light gray studio backdrop instead of white.

Which model made a usable lifestyle scene?

Place the same bottle on a light oak cafe counter next to a glass of iced coffee, soft morning window light from the left, shallow depth of field, clean and uncluttered.

This is where the models looked most alike. Every one built a plausible café, kept the bottle upright, and left the label alone. The differences are framing: GPT Image 2.5 and Qwen Image 3.0 pushed the bottle large and close, Seedream 5.0 Pro set it at the left edge, and two models (Wan 2.7 Image Pro Edit and MAI Image 2.6 Edit) kept the 3:4 shape of the input instead of going square. Some models added milk to the iced coffee, which the prompt never ruled out.

Which model kept the hand and bottle credible?

Show one adult hand holding the same bottle at chest height in front of a plain light gray wall, natural daylight, so the real size of the bottle is clear.

Every model produced one hand and one bottle. At full resolution, all seven gave one hand with five fingers, no extra digits and no merged knuckles. The one real change: MAI Image 2.6 Edit printed light paper stains onto the label that are not in the original photo.

Which details changed across the 21 images?

All seven models kept all three lines of label text in all three shots, so the table below is about the other two things: the background and the framing.

Model	Packshot background	Product height in frame
Nano Banana 2 Edit	70% of edge pixels at 255	82%
Nano Banana Pro Edit	0% at 255 (avg 248)	78%
GPT Image 2.5 Sunburst Edit	0% at 255 (avg 251)	93%
Seedream 5.0 Pro Edit	1% at 255 (avg 253)	79%
Qwen Image 3.0 Edit	0% at 255 (avg 250)	82%
Wan 2.7 Image Pro Edit	0% at 255 (avg 234, gray)	61%
MAI Image 2.6 Edit	0% at 255 (avg 252)	64%

Product height is measured from the top of the cap to the bottom of the contact shadow. "At 255" counts pixels in a 10 px border that are pure white on all three channels.

The label held up better than we expected: 21 of 21 images kept HALDEN, Cold Brew Concentrate, and 500 ml. The things that changed were the things nobody checks until a customer does:

• Color. We sampled the glass just below the shoulder. The input read about RGB 199, 112, 3; Nano Banana 2 Edit's packshot read 8, 2, 0. GPT Image 2.5 (206, 113, 2) and MAI Image 2.6 (201, 112, 3) stayed closest. Our input carries a warm kitchen cast, so with a real product, compare against the bottle in your hand, not against the photo.
• White. Close to white is not white. Most packshots averaged 248 to 253, which looks white on screen and fails a strict 255 check.
• Framing. Two models ignored or lacked a square setting and kept the bottle small in a tall frame.

How can you make AI product photos without a studio?

Shoot a readable source photo

Start with a photo where the label is sharp and the product color is trustworthy.

• Label square to the lens. A label shot at an angle gets "straightened" by the model, and that is where letters drift.
• Text sharp in the input. If you cannot read the small print in your photo, the model will guess it.
• One product, nothing touching it. A spoon leaning on the bottle can end up fused to it.
• Fill most of the frame. The models work from the pixels on the product; more pixels, fewer guesses.
• Shoot the true color in daylight if you can. Our warm kitchen light is the main reason the models disagreed on how amber the glass is.

Use a prompt for each shot type

These are the three prompts from the test, plus one for flat lays. End every one with the "keep the label exactly" sentence above.

Turn this phone photo into a clean e-commerce product photo: the same product alone on a pure white background, soft even studio lighting, a gentle contact shadow, straight front view, centered, filling most of the frame.

Place the same product on [surface] next to [one prop], soft [time of day] light from the left, shallow depth of field, clean and uncluttered.

Show one adult hand holding the same product at chest height in front of a plain light gray wall, natural daylight, so the real size of the product is clear.

Top-down flat lay of the same product on [color] paper with [two related items] arranged around it, even soft light, no shadows across the label.

The flat lay prompt was not part of our test run.

Check the image before publishing

• Background not pure white: run a levels or threshold pass that maps anything above about 245 to 255, then check the corners with an eyedropper. This is a 10 second edit in any photo tool, and every packshot in our test needed it.
• Product too small in frame: crop, but keep the longest side at 1,000 px or more, which is what Amazon needs to switch on its zoom view. MAI Image 2.6 Edit's packshot (880 x 1184, product at 64% of the height) drops to about 890 px tall once cropped to 85%.
• Color drift: rerun with a different model rather than fighting it with prompts. In our test the drift came from the model, not the wording.
• Changed label: if the text moves even one letter, paste your real label artwork back over it. For regulated products such as food, cosmetics, and supplements, do this check every time.

Should you use a product photo app or an image API?

Use an app for occasional edits

For occasional listing photos, an editor such as Photoroom or Claid gives you templates, background removal, and an editing interface. Choose this route when a person will review each image and speed in the editor matters more than model choice.

Use an API when you need repeatable catalog output

An API makes sense when images must flow into a catalog or ad pipeline. The test gives a concrete reason to keep model choice open: Nano Banana 2 made the whitest background but changed the bottle color, while Qwen cost less and GPT Image 2.5 preserved amber better. Route each shot type to the model that passed your product check.

Estimate the cost for 50 products

Three images per SKU, 150 images, at the per-image amounts we were billed, with no retries:

Model	150 images
Qwen Image 3.0 Edit	$4.86
GPT Image 2.5 Sunburst Edit	$6.48
MAI Image 2.6 Edit	$7.05
Nano Banana 2 Edit	$9.18
Wan 2.7 Image Pro Edit	$10.13
Seedream 5.0 Pro Edit	$12.15
Nano Banana Pro Edit	$18.39

Budget for some reruns. In our test every image was usable for its label, but none of the packshots was ready for a marketplace main image without the white cleanup.

Generation time matters at this scale too. Average time per image ranged from about 11 s (Nano Banana 2 Edit) to about 104 s (Seedream 5.0 Pro Edit).

How can you reuse product photos for ads and videos?

Expand a packshot for a banner

Kling Image Expansion widened the Seedream packshot from 1:1 to 16:9 for $0.0224 and kept the studio backdrop and shadow consistent. Watch the size: the result was 1360 x 768, so the bottle ended up with fewer pixels than in the 1024 x 1024 source, and our OCR misread the small print on the expanded version even though it still reads fine by eye. Expand from your highest-resolution shot, and keep the original packshot for anything that needs the label sharp.

Turn an approved image into a product video

A clean packshot or lifestyle image is also the best starting frame for an image to video model. We tested six video models on one product photo in AI Product Video Generator, and for full 15 second ads see AI Commercial Generator.

Run the same checks across a catalog

For a 50-product catalog, start with one difficult SKU: reflective packaging, small type, or curved labels. Generate its packshot and lifestyle scene with two models, then review the label, glass color, background, and crop at full resolution. Use the better model for the matching shot type across the catalog.

Modellix puts the tested image models behind one API key. A script can submit each product image, poll the task, save the result, and record the billed amount. The API documentation gives the request flow; read the selected model page for its image field and size settings. Save finished images promptly because hosted result URLs expire.

At the prices billed in this test, three Qwen edits per product would cost $4.86 for 50 products before retries; the same 150 edits with Nano Banana Pro would cost $18.39. The relevant budget is your approved output rate. Reruns, manual cleanup, and review time sit on top of the model charge.

Try one product photo in Modellix before committing a full SKU list. Use the real label artwork as the final check and keep the approved source image alongside every generated variant.

Frequently asked questions about AI product photography

Which AI model made the best product photo? For packshots, GPT Image 2.5 Sunburst Edit kept the amber color close and filled 93% of the frame, but left only 15 px above the cap. MAI Image 2.6 matched the source color more closely but used just 64% of the frame height. Nano Banana 2 Edit made the whitest background and darkened the bottle.

Can I use an AI product photo as a marketplace main image? Check the current marketplace rules, then inspect the actual file. None of our seven packshots had a fully pure white background. We would clean the background and compare the product color with the physical item before uploading a main image.

Will AI preserve my label? All 21 images in this test kept three short label lines. That result does not cover dense ingredients or curved small print. Inspect each output at full size and use your original label artwork if any letters change.

How much did each edit cost? Our 1K edits cost from $0.0324 to $0.1226 per image, as billed in the test. Current prices and output sizes are on each model page.