AI Product Photography: Models, Costs and Code (2026)

AI product photography is the practice of creating product images with generative image models instead of, or alongside, a camera, lights and a studio. Done well, it turns one clean photo of your product into a white-background catalog shot, a kitchen-counter lifestyle scene and a vertical social ad in a few minutes. Done badly, it produces a bottle with the wrong cap, a label written in invented letters, and a return from a customer who expected the thing in the picture.

This guide covers how the process works, which models suit which job, what each image costs, a complete script that batch-generates scenes from a real packshot, and the mistakes behind most bad results.

What is AI product photography?

At its core, AI product photography uses two kinds of image model. A text-to-image model takes a written description (the prompt) and produces a new image from scratch. An image editing model, sometimes called image-to-image, takes one or more existing images as input plus a prompt, and changes them: a new background, new lighting, a different surface, props added around the product.

A few more terms come up constantly:

• Packshot: a plain, well-lit photo of the product on a neutral background. This is your raw material.
• Reference image: any image you hand the model so it knows what your product (or a style) looks like.
• Lifestyle shot: the product shown in context, such as a mug on a breakfast table or a backpack on a trail.
• Seed: a number that fixes the random starting point of a generation, so the same prompt and seed give similar results again.

The single most important idea: a text-to-image model has never seen your product. If you ask for “a 500 ml amber glass dropper bottle with our logo”, it invents one. That is fine for mood boards and concept work. For anything you intend to sell, start from a real photo and edit it.

Where the technique is strong:

• New backgrounds and scenes without building sets
• Seasonal and campaign variants of the same shot
• Re-framing one image for square, vertical and wide placements
• Concept images for products that are not manufactured yet

Where it still struggles: small printed text, fine mechanical details, true scale, glass and polished metal, and products held in hands. All of these can work, but they need closer checking.

How AI product photography works: two workflows

Workflow 1: generate from a text prompt

You describe the whole image, product included, and the model renders it. This is fast and cheap, and it suits early-stage work: testing packaging directions, pitching a campaign idea, or creating background plates (empty scenes) that you composite your product into later. The weakness is accuracy. Every generation is a slightly different product, so it cannot serve as the image a customer buys from.

Workflow 2: edit a real product photo

You supply a photo of the actual product and describe only what should change. The model keeps the product and rebuilds everything around it. This is the workflow most stores should use, because the item in the image is the item in the box. Some models accept several reference images at once, which helps when the product has a front, back and side that all matter.

Which should you use?

Use text-to-image for ideas and empty scenes. Use editing for every image that will appear on a product page, in an ad, or on a marketplace listing. Most teams end up using both: text-to-image to explore a look quickly, then an edit model to apply that look to the real product.

Best AI image generator for product photography

There is no single best model, because product work covers several different jobs. The table below lists the image models available through Unicorn API, what each is suited to, and the price in credits. On a monthly plan one credit costs $0.01. Prices are per output image, and each request is rounded up to whole credits.

Model	Best product use	Key inputs	Credits per image
Flux 3 Klein Edit Image	Placing a real product into new scenes, up to 4K	prompt, image_urls, resolution, seed, num_images	1.6 (1K), 2.5 (2K), 6.5 (4K)
Flux 3 Klein Text to Image	Concepts and empty background plates	prompt, resolution, seed, num_images	1.6 (1K), 2.5 (2K), 6.5 (4K)
GPT Image 1.5	Packaging with legible text, transparent cut-outs	prompt, up to 4 ref_images, quality, background	3 (low), 4 (medium), 22 (high)
Nano Banana 2 Lite	Fast, cheap 1K edits using several angles	prompt, up to 10 ref_images, 7 aspect ratios	4
Nano Banana 2.1	Higher-resolution edits and generations	prompt, image_input, resolution	4 (1K), 6 (2K), 9 (4K)
Seedream 5.0 Flash Image to Image	Scene edits at one flat price up to 2K	prompt, image_urls, size, aspect_ratio	3.24 (any size)
Seedream 5.0 Flash Text To Image	Concept images at a flat price	prompt, size, aspect_ratio	3.24 (any size)
Seedream 5.0 Flash Layer Decomposition	Splitting an existing shot into layers for editing	prompt, image_url	3.24 (any size)

For lifestyle sets at volume

Flux 3 Klein Edit Image is the cheapest way to put a real product into many scenes, and it exposes a seed field, which helps keep a set visually consistent. At 2K it costs 2.5 credits per image, so four candidates in one request cost exactly 10 credits. Seedream 5.0 Flash Image to Image is a good second opinion at a flat 3.24 credits up to 2K; run the same scene through both and keep the better result.

For packaging text and cut-outs

GPT Image 1.5 is known for following prompts closely and rendering legible text, which matters for boxes, labels and in-image price badges. It also supports a transparent background, which gives you a cut-out you can drop onto any page design. Medium quality at 4 credits is the sensible default; use high quality (22 credits) for final hero images only.

For products with several important sides

Nano Banana 2 Lite accepts up to 10 reference images, so you can show it the front, back, sides and a close-up of the label. More references give the model a better picture of the product, which helps when the output angle differs from your main photo. If you need more resolution than 1K, Nano Banana 2.1 goes to 4K. Our Nano Banana 2 pricing breakdown covers that family in more detail.

The fastest way to decide is to run your own product through three or four of these in the image playground and compare. Product shapes vary so much that your results will tell you more than any general ranking. The full list lives in the model catalog.

How to create product photos with AI, step by step

The process below works with any of the edit models above. The steps that people skip (the first and the sixth) cause most failures in AI product photography.

1. Shoot a clean reference. A recent phone is enough. Use soft, even light (a window on an overcast day works), a plain background, the product filling most of the frame, and a sharp, readable label. Shoot a front view plus two or three other angles.
2. List the outputs you need. Write down every placement and its aspect ratio before generating anything: 1:1 for marketplace and product grids, 9:16 for stories and short video covers, 16:9 or 21:9 for site banners, 3:2 or 2:3 for editorial layouts. Generating to the right ratio beats cropping later.
3. Describe the scene, not the product. The reference image already shows the product. Spend your words on the surface, light, camera angle, props and mood.
4. Tell the model what must not change. Add a plain preservation instruction to every prompt, such as “keep the product, label, colour and proportions exactly as in the reference image; change only the background and lighting”.
5. Generate several candidates. Ask for three or four images per scene. Fix the seed once you find a look you like so the rest of the set matches.
6. Inspect at 100% zoom. Check the label letter by letter, count features like buttons or holes, compare colour against the real product, and look at shadows and reflections for physical sense.
7. Finish and export. Pick winners, re-run the hero shots at higher resolution, and name files so you can trace each one back to its prompt and seed.

A prompt structure that works

[shot type] of the [product] [surface and setting],
[lighting: direction, softness, colour temperature],
[camera: angle, framing, depth of field],
[props, kept few and secondary],
[mood or brand feel].
Keep the [product] exactly as in the reference image:
same shape, colour, label text and proportions.
Change only the background and lighting.

A filled-in example: “Professional product photo of the ceramic coffee dripper on a light oak kitchen counter next to a glass carafe, soft morning window light from the left, eye-level camera with shallow depth of field, a few scattered coffee beans, calm and minimal. Keep the dripper exactly as in the reference image: same shape, glaze, logo and proportions. Change only the background and lighting.”

Lighting words carry more weight than most people expect. “Soft window light from the left” and “single large softbox above” produce very different shadows, and matching the light direction in the new scene to the light in your reference photo makes the result far more believable.

A worked example: one packshot into a lifestyle set (with code)

Say you sell a white ceramic pour-over coffee dripper and want five scenes for the product page and ads. The script below sends your packshot to Flux 3 Klein Edit Image once per scene, asks for four 2K candidates each, polls until the job finishes, and saves every image locally. You will need an API key from your keys page set as UNICORN_API_KEY.

import os
import time
from urllib.parse import urlparse

import requests

API = "https://api.unicornapi.net"
HEADERS = {
    "Authorization": f"Bearer {os.environ['UNICORN_API_KEY']}",
    "Content-Type": "application/json",
}

PACKSHOT = "https://example.com/products/dripper-white.png"

KEEP = (
    "Keep the dripper exactly as in the reference image: same shape, glaze, "
    "colour, logo and proportions. Change only the setting and lighting."
)

SCENES = {
    "kitchen": "on a light oak kitchen counter next to a glass carafe, soft morning window light from the left, shallow depth of field",
    "cafe": "on a dark walnut cafe bar, warm backlight, a few coffee beans in the soft-focus foreground",
    "flatlay": "top-down flat lay on pale linen with a ceramic mug and a stack of paper filters, even diffused daylight",
    "outdoor": "on a wooden camping table beside a lake at golden hour, low sun behind, gentle rim light",
    "studio": "on a seamless warm grey paper backdrop, one large softbox above, soft contact shadow",
}


def run_job(model, job_input, timeout_s=300):
    """Create a job, then poll until it succeeds."""
    r = requests.post(
        f"{API}/v1/jobs",
        headers=HEADERS,
        json={"model": model, "input": job_input},
        timeout=30,
    )
    r.raise_for_status()
    job = r.json()

    deadline = time.time() + timeout_s
    while job["status"] != "succeeded":
        if job["status"] in ("failed", "canceled"):
            raise RuntimeError(f"Job {job['id']} ended as {job['status']}")
        if time.time() > deadline:
            raise TimeoutError(f"Job {job['id']} still {job['status']}")
        time.sleep(3)
        r = requests.get(f"{API}/v1/jobs/{job['id']}", headers=HEADERS, timeout=30)
        r.raise_for_status()
        job = r.json()
    return job


os.makedirs("out", exist_ok=True)

for name, scene in SCENES.items():
    job = run_job(
        "flux-3-klein-edit-image",
        {
            "prompt": f"Professional product photo of the ceramic coffee dripper {scene}. {KEEP}",
            "image_urls": [PACKSHOT],
            "resolution": "2K",
            "aspect_ratio": "1:1",
            "num_images": 4,
            "seed": 42,
        },
    )
    for i, output in enumerate(job["outputs"], start=1):
        url = output["url"]
        ext = os.path.splitext(urlparse(url).path)[1] or ".png"
        img = requests.get(url, timeout=60)
        img.raise_for_status()
        path = f"out/dripper-{name}-{i}{ext}"
        with open(path, "wb") as f:
            f.write(img.content)
        print("saved", path)

Each scene costs 10 credits (four images at 2.5 credits), so the whole run is 50 credits, or $0.50 on a monthly plan. You get 20 candidates and will probably keep five to eight. If you want to confirm the cost before running, the API returns the exact price of any request through POST /v1/models/price; the details are in the API documentation.

Adding a transparent cut-out

For product grids and custom page designs, a transparent PNG of the product is often more useful than a full scene. GPT Image 1.5 supports this directly:

curl -X POST https://api.unicornapi.net/v1/jobs \
  -H "Authorization: Bearer $UNICORN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-image-1.5",
    "input": {
      "prompt": "Clean e-commerce cut-out of the ceramic coffee dripper from the reference image, front three-quarter view, soft even studio light, no background. Keep the shape, glaze and logo exactly as in the reference.",
      "ref_images": ["https://example.com/products/dripper-white.png"],
      "quality": "medium",
      "background": "transparent",
      "aspect_ratio": "1:1"
    }
  }'

# Poll with the id from the response until status is "succeeded",
# then read outputs[0].url
curl https://api.unicornapi.net/v1/jobs/JOB_ID \
  -H "Authorization: Bearer $UNICORN_API_KEY"

That costs 4 credits. If you build store tooling inside an editor, the same models are available to coding agents such as Cursor and Claude Code through MCP; see using the API from agents.

How much does AI product photography cost?

Per image, very little. The real cost is your time spent selecting and checking. Here is what common jobs cost on Unicorn API, on a monthly plan ($0.01 per credit) and with top-up credits ($0.015 per credit, never expiring):

Job	Model and setting	Credits	Plan cost	Top-up cost
One lifestyle image, 1K	Flux 3 Klein Edit Image	2	$0.02	$0.03
Four candidates, 2K, one request	Flux 3 Klein Edit Image	10	$0.10	$0.15
One scene edit, up to 2K	Seedream 5.0 Flash Image to Image	4	$0.04	$0.06
Transparent cut-out, medium quality	GPT Image 1.5	4	$0.04	$0.06
Hero image, 4K	Nano Banana 2.1	9	$0.09	$0.135
Hero image, high quality	GPT Image 1.5	22	$0.22	$0.33

Note the rounding: Flux 3 Klein at 1K is listed at 1.6 credits per image, but a single-image request rounds up to 2 credits. Batching with num_images spreads the rounding, which is one more reason to request several candidates at once.

A full catalog, worked out

Take a store with 20 products, five scenes each, and four 2K candidates per scene. That is 100 requests at 10 credits, or 1,000 credits: $10 on a plan, $15 with top-ups. The Starter plan ($25 a month) includes 2,500 credits, enough for that catalog plus a round of re-generations and a few high-quality hero shots. Plan details are on the pricing page.

Compared with a traditional shoot, you skip studio time, set building and prop sourcing for every variant. You do not skip the need for good reference photos, and you still need someone with an eye for what sells to choose and check the images.

What about free tools?

Plenty of sites offer free AI product photography. Before relying on one for a live store, check three things: the maximum output resolution, whether images are watermarked, and whether the terms allow commercial use. At a few cents per image, paid access is rarely the expensive part of the job.

Common AI product photography mistakes

• Generating the product from text. The result looks like a product, not your product. Always edit from a real photo for anything customers will buy from.
• Judging from thumbnails. Warped letters, an extra handle or a missing screw are invisible at grid size and obvious on a product page. Zoom to 100% on every image you keep.
• Letting the label drift. Small text is where models slip most. If your packaging has fine print, use a model known for text, keep the label large in the frame, or composite your original label back over the result.
• Wrong scale. A model will happily make your 10 cm candle look like a vase. Include scale cues in the prompt (“next to a standard coffee mug”) and check props against real sizes.
• Light that does not match. If your packshot was lit from the left and the new scene’s sun is on the right, the image feels off even when viewers cannot say why. Match light direction in the prompt.
• An inconsistent set. Five scenes with five different colour grades look like five brands. Reuse the seed, keep the lighting phrases consistent, and generate the whole set in one session.
• Over-styling the main listing image. Marketplaces often have strict rules for the primary image, commonly a plain white background with the product filling the frame. Read the rules for each channel and keep lifestyle scenes for secondary images.
• Misrepresenting the product. Changing colour, texture or included accessories, even by accident, invites returns and complaints. The image must show what ships.
• Losing your recipe. Save the prompt, model, seed and settings with every final image. When you add a product next quarter, you can match the set exactly.

When a real photographer is still the better choice

AI product photography is strongest at multiplying good source material, and weakest when there is no good source material to start from. A professional photographer is still worth hiring when:

• Your product is glass, chrome, jewellery or anything where reflections define the look
• Texture is the selling point, such as fabrics, leather or food
• You need a small number of flagship images for a launch, a print campaign or retail packaging
• The product must be shown in use by real people, with accurate fit and scale

A practical split many small brands use: pay for one well-lit reference shoot covering every product from several angles, then use AI to produce the lifestyle scenes, seasonal variants and channel-specific crops. The reference shoot pays off every time you generate a new scene, because better inputs give better outputs from every model.

If you want to try it on your own product today, upload a packshot to the image playground, run the same scene prompt through Flux 3 Klein Edit Image, Seedream 5.0 Flash Image to Image and Nano Banana 2 Lite, and compare at full zoom. Within a few minutes you will know which model handles your product best, and the script above will turn that choice into a full set.