AI product images: how to approve the right variant
Compare references, SKU details and the final crop. Use worked examples and a free worksheet to decide whether to approve, regenerate or keep the product photo.
The scene can change. The product cannot.
Approve an AI product image only when the details that identify the variant match a verified reference — and remain readable in the crop the shopper will see. An attractive background cannot make up for a changed neckline, a different print or an invented pocket.
This workflow is for fashion catalog and content teams. By the end, you will have a decision for each version: approve it for a specific use, regenerate it or keep the verified photograph.
Download the CSV worksheet or open the fillable, printable worksheet. It includes a worked example and blank fields. It works outside Studio.
Provei Studio creates fashion images from product references, a model and visual direction. Its documentation states that generation is probabilistic and requires review before commercial use.[3] This workflow and its worksheets were developed with AI assistance to organize the comparison and decision record. The examples are hypothetical: they are not results from a Studio trial or customer cases.
1. Build a reference for the variant, not the style
Open verified photographs for the SKU, the code that identifies the item in your catalog. Check the variant too: color, print and finish. A similar garment is not an approval reference.
Before reviewing any version, record:
- What must not change: neckline, print design and placement, pockets, fastenings, cuffs, hem and any other details that distinguish the item.
- Where each detail is evidenced: the filename and angle or close-up of the verified photograph.
- Where the image will be used: for example, the main product-page image or a campaign asset; include the format and final crop.
If the reference does not show a decisive detail, obtain another photograph or check with the person responsible for the product. Do not use the generated image to fill that gap. For color, use consistent references; if lighting or the display makes two variants hard to distinguish, put the decision on hold.
2. Use three review states — not an average score
Compare reference and result side by side. Zoom in on the details, then check the file at its publication size and crop. Assign a state to each decisive detail.
| State | Evidence needed | Next action |
|---|---|---|
| Confirmed | The detail is visible and matches the SKU reference. | Passes this criterion; continue with the others. |
| Different | A visible change: altered neckline, added pocket or misplaced print. | Reject the version. Correct the direction or reference and review the next one. |
| Unclear | The detail is covered, illegible or lacks sufficient reference evidence. | Put approval on hold. Determine whether the result lacks visibility or the reference lacks evidence. |
Do not add up points. One different decisive detail means rejection, even if everything else matches. If none is different but one is unclear, the version stays on hold. Only when all are confirmed does it pass visual fidelity.
That is not yet permission to publish. Check composition, readability in the channel and usage rights before the final decision. Provei.AI’s documentation also calls for reviewing rights to the references.[2]
3. Worked example: a hidden neckline is not a wrong neckline
Consider fictional SKU DEMO-V01-AZ: a blue garment with a V-neck, fine vertical stripes and no pocket. The hypothetical reference includes a front photograph and a neckline close-up. The three results below are teaching descriptions, not images produced in Studio.
| Criterion | Hypothetical reference | Version A | Version B | Version C |
|---|---|---|---|---|
| Neckline | V-neck, visible in close-up | Round: different | Covered by hair: unclear | Visible V-neck: confirmed |
| Stripes | Fine and vertical | Match: confirmed | Match: confirmed | Match: confirmed |
| Absent on the front | Added pocket: different | Visible front, no pocket: confirmed | Visible front, no pocket: confirmed | |
| Color variant | Blue in the references | Matches in this scenario: confirmed | Matches in this scenario: confirmed | Matches in this scenario: confirmed |
| Intended use | Main image with neckline and front readable | Reject | On hold | Passes visual fidelity |
A: reject for identity, not taste. The neckline and pocket already show a different garment. A useful correction instruction would be: “Keep the V-neck and pocket-free front shown in reference DEMO-V01-AZ; leave the neckline visible.” This is direction for another attempt, not a guarantee of its result. Recheck every detail afterward.
B: resolve visibility. You do not know whether the neckline was preserved. Request a composition without hair covering it. If the reference close-up were missing, you would need to fix the reference first. “Unclear” describes the available evidence, not a proven defect.
C: check the final use. The described version passes fidelity. But if the store’s crop cuts off the neckline, it is not approved for that main image. Adjust and check the crop; if you cannot keep the details readable, use the verified photograph. Approval must identify the file and the use, not just the SKU.
4. Complete the worksheet before releasing the file
Use one row per detail. The same review ID connects the SKU, variant, reference, version and use. The DEMO rows are filled in for teaching; the TEMPLATE rows are for your work.
- Duplicate the template rows and enter the reference file and reviewed version. Never overwrite an old approval with a new generation.
- Describe what you actually see, not just “OK.” For example: “entire V-neck visible; matches the front close-up.”
- Record the state and next action. Set the version decision only after checking every detail.
- Record the reviewer, date, final crop and rights check. Archive the reference and approved file with the worksheet.
In the HTML version, fill in the fields in your browser and print or save as PDF through its print command. Entries are not sent to a server or saved automatically; print before closing. Import the CSV as UTF-8 with comma-separated fields. The worksheet organizes review: it does not detect errors or approve images by itself.
5. When to regenerate — and when to keep the photograph
- Different detail: make another version using the correct reference and a specific correction. Repeat the entire review; fixing one detail may introduce another change.
- Covered or undersized detail: adjust the pose, framing or crop. An unreadable thumbnail is not evidence of fidelity.
- Insufficient reference: stop generation and obtain the missing evidence.
- Persistent mismatch or exhausted deadline: publish a verified photograph you have rights to use. Agree on an attempt and time limit with the team before starting; there is no universal number.
Do not use this process to infer measurements, fabric composition or physical fit. It assesses what the image shows. If shoppers need to check a finish that the composition does not reveal, keep a verified detail photograph on the page.
What to keep to decide whether the workflow is worthwhile
Record versions reviewed, rejection reasons, review time and the released file. This helps locate bottlenecks: missing references, altered details or unsuitable crops. It does not prove increased sales or fewer returns.
To explore the documented workflow for creating, comparing and reviewing fashion images, see Provei Studio. Bring a variant, its references and the completed worksheet — keep the conversation focused on the job that needs solving.
Sources
Official sources consulted on 2026-10-02 (UTC−03). The rubric and teaching cases are editorial proposals in this guide, not automatic Studio criteria.