The warehouse photo workflow in 12 minutes per SKU.

Total time per SKU

Across every station we land at about 12 minutes of operator time per SKU. That sounds like a lot for one product, but it splits across three different roles and runs in batches — one person realistically covers 150 to 200 SKUs per day.

Stations in detail

1. Warehouse intake — 40 seconds

Enter the SKU code or scan a barcode (5 sec), grab the product or shoot it on the shelf (15 sec), capture one or two angles (15 sec), short category note (5 sec).

Bottleneck: products that sit at the back of the shelf or inside boxes. On average every eighth SKU costs another 30 to 60 seconds of unpacking or shuffling. A trained warehouse worker does 80 to 120 SKUs per hour net.

2. Automated processing — zero operator time

Upload, parsing of the product data, category classification, and image generation run in the background. Roughly 45 seconds of compute per SKU, parallelised across the batch. For 200 SKUs it runs overnight.

3. Review — 25 seconds on Pass, 50 seconds on Fail

The review screen opens with source photo and generated version side by side. Operator looks at the image for 3 to 5 seconds, compares proportions and labelling, clicks Pass or Fail.

On Fail: a short comment. "Background too dark." "Labelling wrong." "Shadow too strong." The comment gets pulled into the prompt rules for regeneration. Regen takes 30 to 60 seconds and comes back as a new candidate.

Fail rate in batch 1 of a new project: 25 to 35%. By batch 3 it's usually under 15%, because the fail comments from earlier rounds keep running as rules.

4. Regen rounds — 90 seconds per fail cycle

The model regenerates with the new rule. Operator reviews again. In 70% of cases one regen round is enough. In 20% it takes two. The remaining 10% are categories that structurally don't work well — see AI product photos — when they work, when they fail.

5. Publish to Shopify or WooCommerce — 6 seconds

Once an SKU is finally approved, the connector runs. REST API call with product data, image upload, category mapping. Typically 4 to 8 seconds per SKU depending on shop latency. Runs sequentially, no manual intervention.

Where you can save time

Barcode scan instead of typed SKUs

Saves 3 to 5 seconds per SKU. Across 700 products that's about 45 minutes.

Set category hints up front

If the warehouse is organised by category, you can set the category once per aisle block and drop it as a per-SKU step. A few seconds per SKU, and lower error rate than fixing it after the fact.

Review in pairs

Two operators on one session — one clicks, the other types comments. Roughly doubles effective review speed because the clicker stays in flow and the typing gets delegated.

Where you can't save time

The generation model itself is the bottleneck at 45 to 60 seconds per image. That's not speed-uppable outside of parallel compute capacity — and it's already running in parallel. For planning: 100 SKUs need one night of compute, 500 SKUs need two nights.

What this means for project planning

• 700 SKUs in 10 working days is realistic — 2 days of intake, 3 days of staggered generation, 4 days of review, 1 day of publish
• 300 SKUs in 5 working days follows the same shape
• Under 100 SKUs rarely justifies the setup cost; stock photography or a half-day studio session is often cheaper
• Over 1,500 SKUs should be split into two waves so the fail lessons from wave 1 push down the fail rate in wave 2