For batch AI product videos, build a manifest that connects each real SKU to its approved source, shot brief, generation attempts, review result, and final file. Approve one still per variant, generate a short clip, then check that the product stays correct throughout the motion. Deliver only the reviewed file recorded against that SKU.
Start with a three-SKU pilot in Oakgen's ecommerce product video workflow. Here, “batch” means a planned production group, with jobs coordinated manually. It doesn't mean a verified one-click catalog import, Shopify sync, CSV ingestion, or automatic publishing feature.
Plan your first three product clips, and download the product-video manifest before starting generation.
The SKU workflow at a glance
| Stage | Record you need | Check before moving on | Stop condition |
|---|---|---|---|
| Map sources | SKU, variant, packaging version, source filename | The photo depicts the exact sellable variant | Missing or conflicting source |
| Approve a still | Approved source or edited still and shot brief | Label, color, shape, hardware, and quantity match | Any product identity error |
| Generate a clip | Model, settings, attempt number, cost | Job uses the source in the recorded row | Wrong input or unsupported shot |
| Review motion | Checkpoint notes plus full playback | Product remains accurate at every visible moment | Drift, occlusion, or invented feature |
| Retry selectively | Failure reason and change for the next attempt | One identifiable cause receives a specific correction | Attempt or spend ceiling reached |
| Deliver | Approved filename, destination, reviewer | File, depicted variant, and product destination agree | Unreviewed file or uncertain mapping |
Open the product-video workflow once the source rows are complete. Oakgen's AI video generator provides the image-to-video generation step; keep the manifest in your own spreadsheet.
Scope and research note
This guide is for merchants and agencies that already have photographs of several physical SKUs and need to avoid wrong-variant delivery. It covers the operational gap between “a clip finished generating” and “the right product page received the right approved clip.”
For broad video formats, model selection, and catalog planning, start with the existing AI video for ecommerce guide. This article concentrates on per-SKU mapping, rejection records, and handoff rather than another catalog-wide cost comparison.
We checked Oakgen's current image-to-video implementation on September 30, 2026. It supports model selection, model-specific input fields, and a job cost estimate. The image-to-video form handles source image fields; options vary by model. The ecommerce landing page directs readers to the video generator rather than documenting a catalog ingestion interface.
We also reviewed Shopify's product-media documentation and Higgsfield's catalog-video guide. Higgsfield describes storing SKU and approval tracking in the brand's own system and cautions readers to confirm exact import methods. A general catalog workflow isn't evidence of a native connector in another product.
Oakgen is our product. The pilot below uses fictional identifiers and planning examples, not real generated clips, observed rejection rates, or measured costs. The downloadable CSV contains planned rows with empty result fields. It is a worksheet, not an integration.
1. Choose a pilot that reveals mapping mistakes
Pick an easy product, a close-looking variant, and a harder product whose label or physical feature could change during motion. Three is enough to exercise the different checks without turning the first review into an inventory project.
Suppose a fictional candle shop has an amber cedar jar, an amber citrus jar, and a larger clear jar. These are planning examples only; supply photographs of your own real variants before generating.
| Illustrative SKU | Why it belongs in the pilot | Detail the reviewer must prove |
|---|---|---|
| NP-CED-180 | Straightforward front-facing package | CEDAR wording, 180 g quantity, amber glass |
| NP-CIT-180 | Similar silhouette; easy to swap accidentally | CITRUS wording and its approved label design |
| NP-CED-300 | Different size and material | 300 g wording, clear glass, actual proportions |
Avoid choosing only the prettiest reference. The close-looking pair helps expose copy-and-paste errors; the larger jar tests whether the same shot plan still suits another shape.
Choose one placement for the pilot, such as the product gallery. Review product video examples if you need to pick a format. A restrained product hero is easier to verify than an unboxing shot that opens packaging you haven't photographed.
2. Map each SKU to its own source
Create the manifest before uploading files. The SKU describes the sellable item; the variant describes the relevant difference, such as scent and weight. Keep a product handle or other destination identifier too, because a family of SKUs may share one product page.
Save stable source names such as NP-CIT-180_front_pack-v2.jpg. A path named bottle-final-new.jpg gives the next operator little help when several near-identical files sit beside it.
Use the downloadable CSV in Excel, Google Sheets, or another spreadsheet. Its main fields are:
| Field | What to put in it |
|---|---|
sku, variant, product_handle | Internal item, exact variant, intended product destination |
packaging_version, source_image | Approved version and stable source location |
shot | A short shot identifier, with the full brief stored alongside it |
model, settings | Selected model and the options actually submitted |
attempt | Sequential attempt number for that SKU and shot |
cost, cost_unit | Actual recorded charge and its unit; leave unknown values blank |
qa_status, review_notes | Workflow status and specific findings |
output_file, reviewer, destination | Final file, approver, and intended placement |
Use one row per generation attempt. Duplicate a planning row when you retry, increment attempt, and record the changed settings or brief. Never overwrite the rejected row with the accepted file.
The template marks its examples planned and leaves cost and output empty. Replace the sample filenames; they aren't supplied assets. Don't turn an unknown charge into zero, and don't type accepted until someone reviews the actual output.
These are suggested statuses for your worksheet, not labels automatically synchronized with Oakgen:
planned -> submitted -> review_pending -> accepted -> delivered
|
-> rejected -> next attempt
-> blocked_source
blocked_source means you need another photograph or approved fact, rather than another generation. A missing view is a source problem.
3. Group shots by product constraints
A shared style doesn't mean every SKU can tolerate the same motion. An opaque box hides different surfaces from a transparent bottle. A reflective device can reveal altered ports as it turns; a pouch can wrinkle across its label.
Group products by the demands your shot creates. Front-facing labeled packages can start with a small camera move while the object stays still. Products with several visible hardware details need sources proving those views. Transparent or reflective products need closer checks for changed contents, tint, and edges.
Write one brief per group, then add each SKU's identity facts. Keep the camera move and framing consistent where the sources support it; avoid forcing a large jar into the exact crop of a small jar.
For a pilot, I would choose one modest push-in and a stationary product. Save hands, opening mechanisms, pours, and full rotations for separate tests with suitable evidence. The product-video prompt guide can help describe the shot, while the product consistency workflow covers broader reference preparation.
4. Approve one still per SKU before motion
Use the real source as the starting still when its scene already works. If you need a background change, edit it and approve the still separately before spending on motion.
Review the same SKU facts you recorded in the manifest: variant wording, material, color, shape, quantity, closures, and visible features. Compare against the source rather than another generated variant. An amber citrus jar can't borrow approval from the amber cedar jar beside it.
Store the approved still under a stable name and update source_image if that still becomes the video input. Retain the original photograph in the project so review doesn't depend on the edited still alone.
Package wording needs its own inspection. Use the logo and label preservation guide when a scene looks convincing but a letter or unit has changed. Do that repair before animation.
Prepare the approved still in Oakgen's image editor if a scene edit is necessary. Returning to video with an inaccurate label only makes the later review harder.
5. Generate and record the actual settings
Open the image-to-video workflow, choose a currently available model that accepts your approved source, and use the duration and output options shown for that model. Record what you submit rather than copying a setting name from a tutorial for another endpoint.
Here is an illustrative shot brief for the fictional cedar candle:
Use the supplied NP-CED-180 approved still as the visual source.
The amber candle jar remains stationary and front-facing on the table.
Use a slow, small camera push-in. Keep the complete jar and label visible.
Preserve the jar proportions, amber glass, lid, label layout, and visible
NORTH PINE / CEDAR / 180 g wording. Keep the scene lighting consistent.
No rotation, opening lid, hands, new objects, added wording, or new badges.
For another SKU, replace the source and identity facts using its truth sheet. Don't ask the prompt to invent that SKU by changing the name of the first product.
Record the selected model, duration, aspect ratio, resolution or quality options when present, and any other supplied controls. Save the full prompt in the project with a version number. If the platform exposes a job identifier, capture it too; otherwise record enough submission context to find the attempt without guessing.
Keep the displayed estimate separate from the recorded charge. An estimate helps decide whether to submit; the charge belongs in the attempt row after submission. Leave unknown values blank until you can verify them.
Run the pilot through the ecommerce product video workflow using your own approved sources. The illustrative prompt doesn't come with a generated clip or a prefilled project.
6. Review the variant throughout the clip
Start with a contact sheet showing the approved source beside the first frame, a middle frame, and the last frame. Add frames where motion, reflections, or occlusion change. Then watch the full clip, because a short defect can fall between the sampled checkpoints.
Compare each output to the source in its own manifest row. Don't let the filename tell you what you think you're seeing. Read the variant name and quantity; inspect the material and the distinctive physical detail.
| QA question | Evidence | Rejection reason to record |
|---|---|---|
| Is it the right variant? | Source alongside visible label or distinguishing feature | variant_mismatch plus exact difference |
| Does the shape stay correct? | Outline and proportions during motion | geometry_drift with first visible timestamp |
| Does the label remain accurate? | Close frames compared with approved wording | label_error with changed characters |
| Does the material stay correct? | Glass tint, transparency, texture, and reflections | material_drift with affected interval |
| Does the intended shot work? | Full playback against the shot brief | motion_or_framing with required correction |
| Does the delivery file work? | Downloaded file played at intended size | export_issue with format or crop problem |
These codes belong to the template, not a built-in detector. Write useful notes beside them. Label changed from CEDAR to CITRUS at 00:03 tells the next operator what to fix; looks weird doesn't.
Review similar variants next to each other after individual checks. This helps catch a repeated source or duplicate output assigned to different rows. Keep accepted separate from delivered, because acceptance doesn't prove that someone uploaded the right file.
7. Retry the cause, within a ceiling
Decide an attempt limit and a spending limit before the batch. For example, you might allow one initial attempt and two revisions per shot; that is a planning policy you choose, not a promise about model success.
If rotation reveals an invented rear panel, remove the rotation or photograph the required view. If the label changes after glare crosses it, try less reflective lighting or less movement. If the wrong source went into the job, correct the mapping before considering any prompt revision.
Add a new manifest row for the retry and describe one deliberate change. Preserve the rejected output for diagnosis, but keep it out of the delivery folder. A visually strong failure shouldn't become the final file because someone later searches by SKU and downloads the first match.
Stop when you hit your ceiling, when the needed evidence is missing, or when repeated attempts fail the same product requirement. Replace the shot with an accurate simpler clip, use the real photograph as a static asset, or obtain a better source.
Calculate production cost from the accepted output backward:
Generation cost per accepted clip =
total recorded cost of all attempts / accepted clips
Full production cost also includes:
source preparation + editing + review + delivery work
Don't divide by all completed jobs if some fail review. Keep units consistent, and use current Oakgen pricing and the selected model's displayed estimate for your own planning rather than borrowing a price from an older article.
8. Deliver a manifest with the files
Create a delivery folder containing accepted files only. Use a filename such as NP-CIT-180_hero-push_pack-v2_approved-v1.mp4, then record that exact name in output_file. Include the depicted variant even when several variants share one store product page.
Add the product destination, intended placement, approver, and delivery date in the handoff. The publishing operator should confirm the file against the row, upload it through the store's supported process, and check playback on the actual page.
For Shopify, the current product-media guide says videos and 3D models can't serve as product variant media. A clip depicting one variant may sit in a product gallery; don't assume changing the variant selector will select its matching video. Verify the theme behavior before delivering a variant-specific claim.
Shopify documents adding product media through its admin or app. Its media types page lists supported video formats and limits. Check those requirements at delivery rather than treating a generator's successful download as a successful store upload.
Use the Shopify product-video workflow for placement planning, or the marketplace product-video guide for a different destination. A marketplace may have its own specifications and review process.
Only mark delivered after the correct file appears at the recorded destination. If your agency ends at file handoff, state that boundary and leave publishing confirmation with the merchant.
Frequently asked questions
Is batch AI product video generation the same as catalog sync?
No. A production batch is a tracked set of video jobs. Catalog sync also needs a verified connection to the store, product updates, media mapping, and publishing behavior. This guide uses manual orchestration and a separate delivery manifest.
Can I create every color variant from one product photo?
Use an approved photo of each real variant. A text instruction to recolor a product doesn't prove the actual finish, stitching, hardware, or packaging of a SKU you haven't supplied.
How many SKUs should I put in the first pilot?
Start with three: an easy product, a similar-looking variant, and a product with a detail likely to drift. That makes source mapping and review problems visible before a larger production batch.
How do I catch a clip assigned to the wrong SKU?
Compare the output with the source file recorded in its manifest row, verify the variant-specific label or physical detail, and check the delivery filename and product destination together.
Can a Shopify product video belong to an individual variant?
Shopify's current help documentation says videos and 3D models can't serve as product variant media. Track the depicted variant internally, but verify how the video displays in the product gallery and theme.
What does a usable catalog clip cost?
Calculate cost per accepted clip using every generation attempt, plus any editing, source preparation, and review costs. A displayed generation estimate covers a job, not the entire production workflow.
Can I upload this CSV to Oakgen or Shopify?
The CSV is a tracking worksheet for your spreadsheet. It isn't an Oakgen ingestion format or a Shopify product import file, and downloading it doesn't create jobs or upload videos.
When should I stop retrying a SKU?
Stop at the attempt or spend ceiling you set before production. Also stop when the shot requires an unseen angle or feature; obtain the right source or simplify the brief before spending again.
Build a three-SKU pilot in Oakgen with one approved still per item and a separate row for every attempt. Scale the process after the manifest, review, and delivery checks work on the pilot.
