E-commerce product images and marketing creatives
E-commerce imagery usually combines three kinds of models: a chat model writes image prompts from product data, an image generation model produces scenes and marketing creatives, and an image editing model swaps backgrounds on real product photos. Image models mostly charge per image. For 500 new products a month with 4 scenes and 2 background swaps each, at live prices in each model’s cheapest group, model costs are about ¥1,002 a month.
Prices updated Nov 12, 2026, 8:17 PM|
01
The business problem
Every new product needs a white-background shot, lifestyle scenes, detail-page images and ad creatives. Photo shoots and retouching are expensive and slow, especially before big sales.
Image models can create scenes from a description or change the background and colors of a real photo. The hard part is keeping the product itself intact: logo, text and shape must match the real thing.
The common approach edits the real white-background photo and lets the model change only the background and mood. Fully generated images suit mood shots and creative tests, and someone checks product details before anything goes live.
02
Recommended models
Scenes and creatives (image generation)
doubao-seedream-5-0-260128ByteDance Doubao¥0.3 per request· Standard groupCreates lifestyle scenes, poster backgrounds and ad concepts from prompts, billed per image, handy for producing several variants to test.
Background swaps and edits (image editing)
nano-banana-proGoogle¥0.4 per request· Standard groupUpload a real product photo and change only the background, lighting or setting. The product stays as it is, which keeps it closer to the real item than pure generation.
Prompts and selling points (chat model)
qwen3.5-plusQweninput ¥1.12 / output ¥6.72 per 1M tokens· China Models groupWrites image prompts, titles and short selling points from product attributes and target buyers. Few calls; see the estimate below for its cost.
Recommended models are picked from the live catalog; prices are each model's cheapest group, updated Nov 12, 2026, 8:17 PM. Click a model to see every group's price.
03
Reference architecture
- 1
Collect inputs
Prepare one clear white-background photo per product, plus its name, material, colors, target buyers and the scene you want to show.
- 2
Write prompts
Have the chat model turn product attributes into image prompts with a consistent style, lighting and composition, and list what must not appear.
- 3
Generate and edit
Generate 3 to 5 variants of mood shots; for main images that must match the product exactly, edit the real photo and swap only the background.
- 4
Store immediately
Image URLs in the response are temporary. Download them right away into your own object storage before review.
- 5
Human review
Check shape, logo, text and colors against the real product and confirm you have the right to use the material before listing or advertising.
04
Cost estimate
Assumptions: 500 new products a month; for each, one set of prompts from the chat model (about 600 input and 400 output tokens), 4 generated scenes and 2 background swaps on the real photo, so 2,000 generations and 1,000 edits.
| Item | Model | Usage assumed | Cost |
|---|---|---|---|
| Writing image prompts | qwen3.5-plus | 500 × (600 input + 400 output tokens) | ¥1.68 |
| Generating scenes | doubao-seedream-5-0-260128 | 2,000 requests | ¥600 |
| Background swaps and edits | nano-banana-pro | 1,000 requests | ¥400 |
| Total (per month) | ¥1,002 | ||
Calculated from live prices in each model's cheapest group, updated Nov 12, 2026, 8:17 PM.
Model usage only: storage and human review are not included. Image models mostly charge per image, and resolutions and groups of the same model are priced differently; the table uses the cheapest group. Count the regenerations you will need as well.
05
Sample code
Generate a scene from text, then swap the background of a white-background product photo. Results arrive as a temporary URL or as base64; save them immediately.
import base64
import os
import requests
from openai import OpenAI
client = OpenAI(base_url="https://open.tutujin.com/v1", api_key=os.environ["TUTUJIN_API_KEY"])
def save(image, path):
"""Results come back as a temporary URL or as base64. Store the file right away."""
if image.b64_json:
data = base64.b64decode(image.b64_json)
else:
# some image hosts reject requests without a browser-like User-Agent
r = requests.get(image.url, headers={"User-Agent": "Mozilla/5.0"}, timeout=120)
r.raise_for_status()
data = r.content
with open(path, "wb") as f:
f.write(data)
# 1. a lifestyle scene from a text prompt (prompts can be written by qwen3.5-plus)
scene = client.images.generate(
model="doubao-seedream-5-0-260128",
prompt="A white ceramic mug on a wooden breakfast table, soft morning light, product photography",
)
save(scene.data[0], "mug-scene.png")
# 2. keep the real product, change only the background
with open("mug-white-background.png", "rb") as f:
edited = client.images.edit(
model="nano-banana-pro",
image=f,
prompt="Keep the mug exactly as it is and place it on a marble kitchen counter",
)
save(edited.data[0], "mug-marble.png")FAQ
What does imagery for 500 new products a month cost?
With 4 scenes and 2 background swaps per product plus one set of prompts, the cheapest groups come to about ¥1,002 a month. Image models charge per image, so more regenerations mean a higher bill.
Can generated product images differ from the real product?
Yes, especially with pure text-to-image, where logos, text and details can change. For main images that must match, edit the real photo, change only the background and check the result by eye.
Do the returned image URLs last?
No. They are temporary and expire after a while, so download and store the files as soon as you receive them.
Can I use generated images commercially?
Whether you can, and whether someone else’s trademark or likeness is involved, depends on the model provider’s terms and on where your uploaded material came from; confirm this yourself. We do not judge copyright ownership of generated content.
How do I choose between image models?
Run the same prompts on a few models, compare results and per-image prices, then pick your main model. For edits of real photos, choose a model that supports the image editing endpoint.