Cutout batches rarely fail at the model. They fail in the folder you hand it: mismatched framing, stray alpha channels, filenames like IMG_4471 copy 2.jpg, and a 40 MB master that never had a chance of clearing an upload check. The work that happens before submission — normalising, naming, sampling and costing — decides whether a run finishes cleanly or turns into a week of manual edge repair.
What a Cutout Model Actually Needs From Your Product Shots
Start with the constraints the service publishes, because they shape every decision upstream. The BiRefNet general model on Y-PAD accepts png, jpeg and webp files up to 25 MB each, and by default returns a PNG cutout with an alpha channel; passing output=mask returns the mask alone. That default matters for prep: if your compositing step expects premultiplied alpha or a fixed canvas size, you are responsible for that transformation after download, not the model.
Resolution is the lever most teams get wrong in both directions. Pricing runs at $0.0005 per image up to 1 MP, pro rata above that, capped at 4 MP — so a 30 MP studio file does not buy a better edge, it just costs more per item and moves more slowly through storage. Pull a handful of representative shots, cut them at full size and again downsampled to roughly 1 MP, composite both on white and on mid-grey, and pick the size where the edge quality you need actually holds. For catalogue thumbnails that is often the 1 MP version; for zoom-enabled hero images it may not be.
Source alpha is the second trap. Shots exported from an earlier retouch may already carry transparency, and images pulled from a marketplace feed sometimes arrive with a soft vignette or a baked-in white background. Flatten anything with an existing alpha channel onto a neutral background before the batch, and note which items were already knocked out — those are the ones most likely to come back with a nibbled edge, and they deserve a review flag rather than silent inclusion.
Then look at framing and the awkward categories. The model needs some background around the subject; products that touch or cross the frame edge give it nothing to reason about on that side, so recrop with a consistent margin. Glass, brushed metal, sheer fabric and anything with a deliberate cast shadow are the predictable hard cases — not because the model cannot handle them, but because their edges are genuinely ambiguous and a human needs to sign off. Build that list now, while the product owner is still available to answer questions.
Building the Prep Pipeline: Manifests, Limits and Polling
Drive the batch from a manifest, not from a directory listing. A CSV or JSON file with one row per SKU — source key, target filename, desired output size, review flag, category — gives you deterministic names, so a re-run writes to the same key instead of producing cutout_final_v3 variants nobody can reconcile later. It also makes partial retries trivial when three items out of two thousand come back wrong.
Normalise before you upload. Convert everything to png, jpeg or webp, confirm each file sits under 25 MB, and strip colour profiles down to sRGB so products photographed in different sessions do not drift in tone once composited onto a shared page background. Keep the originals; route batch output into a separate prefix so the prep work stays reproducible.
Respect the batch shape. The API accepts up to 100 files per call and up to 2,000 items per batch, so a large catalogue is naturally chunked — and chunking by category rather than folder order contains the damage when one product line behaves badly. Upload links expire after one hour, so stage files close to submission time instead of pre-uploading overnight, and download results promptly: outputs are retained for seven days. There are no webhooks, so your poller needs backoff, idempotent state tracking and a record of batch and item IDs. Expect a cold start of 10 to 30 seconds, occasionally minutes, with later items in a batch riding on the same start — an argument for fewer, larger batches over a stream of tiny ones. If you would rather script the run than click through a UI, a background removal api with Replicate-shaped JSON keeps request-building familiar, and trying a few files in the playground first catches schema mistakes before you queue a thousand items.
One practical snag: uploads return 402 until the account’s first payment clears, because there is no free storage. Fund the account and store the API key (shown once, op_live_ prefix) before the pipeline’s first real run, not halfway through it.
Cost, Holds and QA Before You Scale
The cutout fare is easy to reason about: $0.0005 per image up to 1 MP, pro rata above that up to 4 MP, with 1,000 cutouts working out to $0.50 as of 7 November 2026. Every request gets batch pricing with no minimum charge, so a single test file costs the same per unit as the thousandth. Only successful items are charged; failed and cancelled items cost nothing. Note that BiRefNet’s list and launch prices are identical, so a launch seat changes nothing on this model — budget cutouts at the one published number rather than shopping for a discount tier.
Cash flow still needs planning. At submit, the platform holds the sum of maximum units multiplied by list price, then settles succeeded units at list (or launch price, where a seat applies) and returns the unused hold. Packs start at $10, with $25 and $100 tiers, and the first payment cannot exceed $25; for seven days after that first payment, the day’s charges plus open holds cannot pass $25. That cap is worth modelling before you queue a very large batch, even a cheap one. Prices are provisional beta fares published for 7 November 2026, so confirm them on the live pricing page before you commit a budget.
A pre-submit review loop that scales
Run a stratified sample rather than a random one: a few items from each lighting setup, each category, and every flagged already-transparent item. Compare cutouts composited on white and on a dark surface, because halos that vanish on white show up badly on a dark product page. Use output=mask on selected files to inspect the matte itself — thin straps, cable loops and mesh are where a mask review pays for itself. Then lock that sample set and keep it as a regression suite: every future batch gets the same dozen images, and any drift in edge behaviour surfaces before it reaches twenty thousand SKUs.
Keep the review trail with the batch: flagged SKUs, re-run reasons, mask snapshots for problem items, and the manifest version that produced them. When the next seasonal shoot lands with different lighting, that record is what tells you whether the model drifted or the photography did.
Brooklyn-born astrophotographer currently broadcasting from a solar-powered cabin in Patagonia. Rye dissects everything from exoplanet discoveries and blockchain art markets to backcountry coffee science—delivering each piece with the cadence of a late-night FM host. Between deadlines he treks glacier fields with a homemade radio telescope strapped to his backpack, samples regional folk guitars for ambient soundscapes, and keeps a running spreadsheet that ranks meteor showers by emotional impact. His mantra: “The universe is open-source—so share your pull requests.”