Practical guide
AI Product Photography Workflow: From Prompt to Final Asset
A strong AI product image starts with fixed product facts, a simple composition, one lighting direction, and a final manual check for labels and geometry. Draft cheap at 1K, batch variants by changing one variable per run, and spend on 2K or 4K only after a prompt survives review. Run the loop the same way every time and it becomes a repeatable pipeline instead of one-off luck, with per-image credit costs you can read before you generate.
By Byron Johnsie · Spatial designer
Last updated:
Step-by-step
- 01
Fix the product facts first
Write down what must never change - bottle shape, cap, label hierarchy, colours, materials - and reuse that line verbatim in every prompt. Anything you do not explicitly protect is a detail the model may invent or let drift. Treat this list as the contract that every later step checks against.
- 02
Choose the reference image
Upload the sharpest photo of the product you have: fully visible, evenly lit, no clutter. Kansova uses one reference image per generation, so this single photo carries the product's identity through the whole workflow. For reference edits, the prompt should describe only what changes - environment, light, surface - never the product itself.
- 03
Set one lighting direction
Decide the light before the props: soft morning window light, one hard side light, or even overcast light. Name a single direction in the prompt, because two competing light sources produce muddy shadows and washed-out labels. If a run disappoints, change the light before you change anything else.
- 04
Build the prompt in a fixed order
Use the same structure every time: preservation line, environment, lighting direction, framing (top-down, three-quarter, eye level), and aspect ratio. Change exactly one variable per run - background, light, or props - so you always know what caused a difference. A prompt that asks for everything at once gives you nothing to compare.
- 05
Draft at 1K and batch the variants
Generate two or three candidates per direction at 1K, where GPT Image 2 costs 2 credits per image, then batch variants by swapping one scene or light word at a time. The free plan's 100 starter credits cover about fifty 1K/2K GPT Image 2 drafts, so exploration is cheap. Keep every run on the same prompt skeleton so the batch stays consistent.
- 06
Review every output against the original
Zoom in and compare labels, logos, small printed copy, edges, and geometry with the source photo. AI preserves products on a best-effort basis and small text drifts first, so plan to replace drifted copy in design software rather than re-rolling forever. Reject anything you cannot verify against the real product.
- 07
Re-run the winners at final size
Only approved prompts deserve final resolution. GPT Image 2 and Nano Banana cost the same at 1K and 2K - 2 and 4 credits per image - while 4K doubles the base to 4 and 8 credits. On GPT Image 2.5 the quality ladder runs 2/3/6/10/22 credits at 1K from low to max quality, with 4K at four times the base.
- 08
Export, name, and archive
Name files by SKU, scene, and version - for example sku-1042-oak-table-v2 - so variants stay traceable as the campaign grows. Download the finals from the asset library, and keep the original product photo and the approved prompt on file so any asset can be regenerated later. Paid plans include a commercial use license, subject to the Terms of Service.
Comparison data
| Stage / Model | Input or size | Output / credits | Review gate / note |
|---|---|---|---|
| Brief | Product facts | Prompt constraints | No invented claims |
| Concept | Text prompt | 1K draft | Composition |
| Refine | Reference image | Edited draft | Product fidelity |
| Batch variants | Same prompt, one changed variable | 3-5 candidates | Consistency across the set |
| Final | Approved prompt | 2K/4K image | Labels and edges |
| Deliver | Final file | Named campaign asset | Usage rights and naming |
| GPT Image 2 | Per image, 1K/2K | 2 credits | 4K costs 4 credits |
| Nano Banana | Per image, 1K/2K | 4 credits | 4K costs 8 credits |
| Nano Banana Pro | Per image, 1K/2K | 8 credits | Reference-guided edits |
| Seedream 5 Lite / Pro | Per image, 1K | 3 / 4 credits | Pro caps at 2K (6 credits) |
| GPT Image 2.5 | Per image, 1K by quality | 2/3/6/10/22 (low to max) | 4K is 4x the base cost |
Recommendations by scenario
Ecommerce SKU batch
Run one shared scene brief across the catalog: white-background pack shot first at 2 credits per image on GPT Image 2, then the same styled-scene prompt for each SKU. On the Basic plan, 300 monthly credits cover 150 GPT Image 2 images at 1K/2K.
Brand campaign direction
Explore with text-to-image at 1K to pick a direction, then replace invented product details with a reference-guided pass using the real pack shot. Spend on 4K only for the approved hero image.
Social media variants
Keep the prompt skeleton and change only the framing line per run: 1:1 for the marketplace, 4:5 for feed posts, 9:16 for stories. One variable per run keeps the set recognisably the same product.
Architecture and interior context
For furniture and fixtures, describe the space and its light separately from the fixed product constraints, and reuse one interior style brief across the set so rooms and renders stay consistent.
Marketplace compliance refresh
When listing rules change, regenerate compliant white-background shots through the same prompt for every SKU, then verify each label against the original photo before uploading.
Concrete examples
- 01Keep the bottle shape, cap, and label hierarchy unchanged. Place it on pale blue glass with soft water caustics and a clean horizon.
- 02Create a top-down campaign still of this watch on brushed steel, one hard side light, precise shadow, no extra objects.
- 03Preserve the chair design. Move it into a minimal concrete gallery with late-afternoon light and neutral styling.
- 04Keep the jar, lid, and label exactly as shown. Set it on a warm oak shelf with soft morning window light from the left, shallow depth of field, no props.
- 05Same product, same label. Light grey studio sweep, soft even light, soft contact shadow, centered composition, 1:1 for the marketplace.
- 06Preserve the bottle and cap exactly. Top-down on matte ceramic tile, one hard side light from the right, a long precise shadow, negative space on the left for text.
Limitations and pitfalls
- AI output is not proof of physical product dimensions.
- Small printed copy may need to be replaced in design software.
- Reference editing can unintentionally change unprotected details.
- Kansova uses one reference image per generation, so multi-angle sets need one generation per angle.
- Results vary between runs; generate two or three candidates per direction and keep the best.
- Model availability, resolutions, and reference support can vary by upstream provider.
Continue in Kansova
Nano Banana guide
Reference-edit workflow details.
OpenGPT Image 2 guide
Detailed instruction workflow.
OpenAI Image Generator
Start an image task.
OpenAI Product Photo Generator
The product-specific tool guide with parameters and credit costs.
OpenProduct photos without a studio
The marketplace-first version of this workflow, from one photo to final assets.
OpenHow AI generation credits work
The full credit cost table for every image and video model.
OpenFrequently asked questions
Do I need a real product photo to start?+
No. Text-to-image is enough to establish direction, then replace invented details with a reference-guided pass using the actual pack shot.
Which resolution should I use for drafts versus finals?+
Draft at 1K to check composition, then produce the final at 2K or 4K once the prompt is approved. On GPT Image 2 and Nano Banana, 2K costs the same as 1K - 2 and 4 credits per image - while 4K doubles the base to 4 and 8 credits.
How do I stop the model from changing the product?+
State the fixed facts explicitly (shape, cap, label hierarchy, colour) and upload a reference image. Unprotected details can still drift, so check them before delivery.
How much does a typical workflow run cost?+
Draft cheap, finish small. Ten 1K/2K drafts on GPT Image 2 cost 20 credits, ten Nano Banana reference edits cost 40 credits, and a single 4K final is 4 credits on GPT Image 2 or 8 on Nano Banana. The free plan's 100 starter credits cover about fifty GPT Image 2 drafts, and the exact total is shown before every generation.
How many images do the monthly plans include?+
It depends on the model, because one credit pool covers all of them. Basic ($9.90/month, 300 credits) covers 150 GPT Image 2 images or 75 Nano Banana edits at 1K/2K. Pro ($19/month, 750 credits) covers 375 GPT Image 2 images. Studio ($39/month, 1,700 credits) covers a mixed workflow of drafts, finals, and selected video clips.
Can an AI image prove product dimensions?+
No. The output is a visual concept, not evidence of physical size or specification.
Can I use the images commercially?+
Free accounts are for personal use. Paid plans include a commercial use license subject to the Terms of Service.
Sources and methodology
The references below support product capabilities and workflow guidance. Recommendations are editorial judgements about task fit, not results from a controlled quality test. Current Kansova settings and credit rules are documented on the product facts page.
- EvoLink documentationProvider capability and parameter reference.
- Kansova factsCurrent product capabilities and limitations.
- Kansova pricingCurrent plans and credit allowances.