Upscaling looks like the simplest job on any image API: send a picture, get a bigger picture back. Then you multiply it by a catalogue of 40,000 product shots and the arithmetic stops being trivial. A 4x upscale quadruples width and height, so the output carries roughly sixteen times the pixels of the input — which has real consequences for storage, download time and format choice, even when the per-image fare looks like rounding error. This guide walks through the three things that actually decide what a bulk 4x job costs and how it behaves: input megapixels, the pro rata billing model, and the file types on both ends of the pipeline.
Why Input Megapixels, Not Output, Set the Bill
Most upscalers quote per image, which is fine for a demo and awkward at scale. A 512×512 avatar and a 1600×1200 photo take wildly different amounts of compute, so a flat price either overcharges the small files or loses money on the big ones. Pro rata pricing fixes that by keying the fare to input size. On Y-PAD’s Real-ESRGAN x4plus model, the published rate as of 7 Nov 2026 is $0.0010 per image up to 1 MP input, pro rata above that, up to 2 MP. The scale parameter accepts 2 or 4 and defaults to 4.
In practice that means your cost estimate is a multiplication, not a lookup. A 900×900 JPEG is about 0.81 MP, so it bills near $0.00081. A 1,024×1,024 file is about 1.05 MP and lands just over a tenth of a cent. Ten thousand images averaging 1.4 MP works out to roughly $14.00 — and because the rate is proportional rather than tiered, averaging your corpus is mathematically safe in a way it never is with stepped pricing. You can see the exact fare breakdown on the image upscale api model page before you commit a batch.
The output side is where teams get surprised. A 4x run on a 1 MP input returns roughly a 4,096×4,096 image — around 16.8 MP, before any alpha handling. That is sixteen times the bytes in a raw buffer, and with PNG as the default output format, sixteen-ish times the disk. If you are upscaling 20,000 images, you are not planning a 20,000-file download; you are planning a multi-hundred-gigabyte one. Decide early whether you need the full 4x output archived, or whether the resized derivatives are the real deliverable.
Profile the corpus before you submit anything. You do not need to decode images to get dimensions — most libraries expose a header read that costs almost nothing. Bucket your files by megapixel band, then price each bucket at the published rate. That single script turns a vague “what will this cost” into a number you can defend, and it flags the outliers: a handful of 6,000-pixel-wide originals can dominate a batch that otherwise looks uniform. Note also that the published pro rata band tops out at 2 MP input, so anything larger needs checking against the live pricing page rather than extrapolation.
Pro Rata Billing Mechanics: Holds, Settlement and Launch Seats
Y-PAD bills batches rather than individual requests: every request gets batch pricing with no minimum charge, so one file costs the same per unit as a thousand. The mechanics matter when you run large jobs. At submit, Y-PAD holds sum of max_units × list price against your balance. At settlement you pay for succeeded units only — at launch price if you hold a launch seat, otherwise at list — and any unused hold returns to your balance. Failed and cancelled items cost nothing, and the prices used are stored with the batch, so a later price change cannot retroactively alter a job you already ran.
Two details change how you should size a first run. First, an image with an alpha channel takes a second pass and costs the same again. If your catalogue is full of transparent PNGs — cut-out products, logos, UI assets — your effective rate is double the headline number, because each RGBA file is processed twice. Filter by mode during preprocessing and price those files at 2× before you promise anyone a budget. Second, for seven days after your first payment, that day’s charges and open holds together cannot exceed $25, and the first payment itself can’t exceed that figure. A first batch of 30,000 images at 1.5 MP would demand a $45 hold and get blocked by the cap, not by your balance. Stage the rollout: submit, poll, let it settle, then ramp.
Launch seats deserve a note here because they are frequently misread. The $25 Standard pack and the $100 Bulk pack may claim a launch seat while seats remain — 40 in total, one per account and per card, assigned at payment — giving the first $100 of usage at launch price. But that discount applies to a specific list: Whisper large-v3, Qwen3-VL 8B, GLM-OCR, DeepSeek V4.1 Flash, Seedance 2.5 I2V/T2V, Pruna P-Video and WAN 2.6 I2V. Real-ESRGAN is not on it. Its list and launch price are identical at $0.0010, so a seat neither helps nor hurts your upscale bill. All figures here are provisional beta fares as of 7 Nov 2026 — confirm them on the live pricing page before you forecast a quarter.
File Types, Size Limits and What a 4x Upscale Returns
The input contract for Real-ESRGAN is narrow and worth memorising: png, jpeg and webp, largest file 25 MB. No TIFF, no 16-bit PSDs, no HEIC straight off a phone. If your source pipeline produces any of those, convert upstream — and convert to something lossless where you can afford it, because a 4x upscale amplifies JPEG ringing and blocking along with everything else. Feeding the least-compressed original you have is the single cheapest quality decision in the whole job. Re-encoding a JPEG to PNG does not recover detail; it just stops you losing more.
Outputs are png, jpeg or webp, with PNG as the default. That default is the right choice when you need to preserve transparency or feed a downstream matting or compositing step, and the wrong one when you are generating browse-size derivatives for a storefront. A 16.8 MP PNG per product is a lot of bytes to serve; a webp at the same dimensions is far lighter. Set the format explicitly per job rather than accepting the default out of habit, and remember that if the input carried alpha, the second pass matters for the output too — keep a format that can hold the channel.
Operationally, plan around a few fixed constraints. Upload links last one hour, so presigned URLs are for immediate submission, not for queuing overnight. You can send up to 100 files per call and up to 2,000 items per batch. Outputs stay available for seven days, which means download and persist is part of the job, not an afterthought. There are no webhooks — you poll — and cold starts run 10 to 30 seconds, occasionally minutes, with later items in a batch riding the same start, so splitting a run into many tiny batches usually makes wall-clock time worse rather than better. Uploads also return 402 until your first payment, since there is no free storage tier. Responses are JSON shaped like Replicate’s, so existing clients port with light edits, and the playground lets you try a couple of representative files before you write any code.
Baghdad-born medical doctor now based in Reykjavík, Zainab explores telehealth policy, Iraqi street-food nostalgia, and glacier-hiking safety tips. She crochets arterial diagrams for med students, plays oud covers of indie hits, and always packs cardamom pods with her stethoscope.