Fernway Home is a fictional three-person online shop selling mugs, throws and lamps. Its last half-day photo shoot covered 12 products and cost $1,400 once the photographer, a rented kitchen and props were paid. Every new colourway means another shoot or another round of retouching. So the owner asked a simple question: could AI image generation produce usable lifestyle shots and marketplace images, and at what real cost per usable picture?
We ran the test over one morning. Twenty-seven images, three model tiers, $0.61 in total API charges and no failed jobs. Below is what held up and what did not.
The setup
We used the Seedream 5.0 family, which comes in three tiers that take the same request format. Pro is the quality-first option at $0.0293 per 1K image. Flash is the budget tier at $0.0137 per image at any size. Lite outputs 2K to 4K at $0.028 per image and can return up to 15 images in one request. All three use one asynchronous endpoint: you submit a job, get a task ID back, then poll until the image is ready. A request looks like this in Python:
payload = {
"model": "seedream-5-0-flash", # or seedream-5-0-pro / seedream-5-0-lite
"prompt": "Lifestyle product photo ... No text, no logos, no people.",
"size": "4:3",
"resolution": "1K" # Lite accepts 2K, 3K or 4K only
}
r = requests.post(API_BASE + "/v1/images/generations", headers=HEADERS, json=payload)
task_id = r.json()["data"][0]["task_id"]
# poll GET /v1/tasks/{task_id} until status == "completed"
Each product had a written spec, the same kind a buyer would see on the listing. The mug, for example: about 9 cm tall, matte oatmeal glaze with fine brown speckles, an unglazed sand-coloured foot ring, a rounded handle on the right. We scored each image against that spec, not against how attractive it looked.
Test 1: does the product match the listing?
All three tiers got the obvious details right: speckles, unglazed foot, handle on the right, napkin, oak table. The difference was shape. Pro produced a plain cylinder in all three runs. Lite produced a rounder, slightly tapered mug, but it did so consistently across five images, including a batch of four returned by a single request. Flash was fast, around 10 seconds per image, yet four of its five mugs came back faceted or dented, a "hand-thrown" look the spec never asked for.
The Seedream API bills per delivered image, and failed tasks are not charged, so rejects are what change the economics. Flash is the cheapest per image, but with one usable mug out of five its cost per usable image was about $0.069. Pro came to $0.029 and Lite to $0.028, because nothing had to be thrown away.
Test 2: colour accuracy
Returns over "the colour looked different online" are expensive, so we gave the throw an exact colour, sage green #9CAF88, and measured the median colour of all blanket pixels in each image. We used Delta E, the standard colour-difference score, where values under about 5 are hard for most shoppers to notice. Pro scored 4.4. Flash scored 18.9 and Lite 26.3; both rendered a darker, greyer olive. Some of that is the lighting each model chose, but for a listing, the darker version would simply be wrong.
Test 3: text in the image
Shops often want a price card or size label in a lifestyle shot. We asked for a card reading exactly "Ribbed lamp – 42 cm – $68". Pro and Flash spelled it perfectly. Lite also got every word right but dropped the first dash and broke the line. For a price, close is not good enough, so we would still add text in an editor and keep the generated image clean.
Test 4: marketplace white backgrounds and edits
Marketplaces usually want the main image on pure white. None of the three tiers delivered that from a text prompt alone. We checked the outer border of each packshot: only 21.5% of Pro’s border pixels reached a brightness of 250 out of 255, Lite reached 49.4% and Flash 94.4%. The fix was cheap. Sending Pro’s packshot back to Flash as a reference image with a "pure white background" instruction took 6 to 7 seconds and $0.014, and lifted the border to 98.7% and 99.8% in two runs. Only 0.1% of the mug’s pixels changed noticeably.
We then asked each tier to place the same mug on a windowsill and keep it unchanged. All three kept the shape and the handle. Pro held the glaze hue most closely, with a colour shift of 1.4, while Flash and Lite warmed it to about 7. Pro was also the slowest here, at 86 seconds.
What it adds up to
Across the 27 images, median generation time was 44 seconds for Pro, 10 for Flash and 30 for Lite, with a four-image Lite batch finishing in 68 seconds. For Fernway, the sensible split is Pro for hero lifestyle shots where colour and shape must be right, Flash for quick clean-up edits such as white backgrounds, and Lite when it needs several 2K or larger variants at once. Even at ten attempts per final image, a 12-product catalogue would cost under $5 in API fees, against $1,400 for the shoot. Staff time for checking and retouching is the real cost, and it should be budgeted.
Do not mislead the buyer
The cheapest image is worthless if it causes a return or a complaint. Our rules: generate only from an accurate written spec, compare every image side by side with the real product before publishing, never show a colour, size or finish the item does not have, and label lifestyle scenes as illustrative where a marketplace or local rule asks for it. AI images should show the product in context; they should not invent a better product than the one in the box.