Real cost of AI watermark removal — browser vs SaaS vs self-hosted GPU

All numbers come from public pricing pages, measured runtimes, and the US average industrial electricity rate of $0.16/kWh as of mid-2026. Reproducible from the math in this article.

If you only have time to read three lines: a Telea-based browser tool costs roughly $0.0000003 per image (the electricity to power your phone for a fifth of a second). A typical SaaS watermark tool at $9.99/mo costs $0.005 per image if you use it 2,000 times a month, or $0.05 if you only use it 200 times. Self-hosting LaMa on an A100 at $1.50/hr costs $0.00008 per image in GPU time alone, plus bandwidth, plus the 50 MB weight download that every user pays for once.

First-screen answer: For low-to-medium volume (under ~50,000 images per month) the browser tool is 4-5 orders of magnitude cheaper than SaaS per image, and 2-3 orders of magnitude cheaper than self-hosted LaMa. The crossover happens around 50,000 images per month if you actually run the GPU 24/7. Until then the SaaS subscription is mostly paying for the website and the support inbox, not the compute.

Three approaches, three cost stacks

When you remove a watermark you pay for something. What you pay for and who you pay depends on where the work happens:

  1. Browser tool: the work runs in your browser on your phone or laptop. You pay for the electricity to power your device, and nothing else.
  2. SaaS subscription: you upload your image to a vendor’s server. The vendor runs the model. You pay a monthly fee that covers their compute, their bandwidth, their storage, their support inbox, and their profit.
  3. Self-hosted GPU: you rent a virtual GPU from a cloud provider, run the model yourself, and charge yourself the GPU-hour rate. You pay for the GPU time, the egress bandwidth, and the disk to hold the model weights.

Same image, same model underneath (sometimes), three completely different price tags. We break each one down below.

1. The browser tool: electricity only

The site you are reading runs Telea 2004 in your browser. The full pipeline (decode + inpaint + encode) for an 800x600 image on a 2024 mid-range phone takes about 250 ms. Phone SoC power draw under sustained compute is roughly 4-6 W; we will use 5 W as the midpoint.

Per image:

At any human scale of usage — a thousand images, a hundred thousand images, a million images — the electricity is rounding noise in your monthly phone bill. The actual cost to you is zero; the actual cost to the grid is one ten-thousandth of a cent. The operator of a browser tool pays nothing per image: bandwidth is the page itself, compute is the user’s device.

Source: US Energy Information Administration Average Retail Price of Electricity, residential sector, July 2026 release.

2. The SaaS subscription: what you actually pay for

The market has settled on roughly three price tiers for watermark-removal SaaS. Names vary; numbers do not. We use ranges we have seen across the segment in 2026:

TierMonthly priceTypical image capPer image if you max itPer image if you use 200/mo
Starter$4.99500 images$0.010$0.025
Pro$9.992,000 images$0.005$0.050
Business$29.9910,000 images$0.003$0.150
Pay-per-image——$0.10 - $0.50 per image—

That per-image column is what matters and most users miss it. A $9.99/mo subscription is not “unlimited watermark removal”; it is “$9.99 for up to 2,000 images, and $0.005 per image if you use it fully, and $0.05 per image if you only use 200 of them”. The break-even for the Business tier is roughly 600 images per month. Below that, you are overpaying.

What the SaaS price covers that the browser tool does not:

So the per-image cost of a SaaS is dominated by shared costs (support, profit) spread across active users, not by the per-image GPU time. If you max your tier, you are getting closer to a fair price. If you do not, you are subsidising the heavy users.

3. The self-hosted GPU: paying for the machine, not the image

You rent a virtual GPU. You run the model yourself. You pay the cloud. This is the model a serious image-processing shop picks when they are handling 100k+ images per month or have compliance reasons to keep data in-house.

Per-image compute cost on common GPU options in 2026 (spot pricing, us-east-1 equivalents):

GPUHourly rate (spot)Time per image (LaMa)Per-image GPU
NVIDIA A100 40GB$1.50/hr200 ms$0.0000833
NVIDIA A10G 24GB$0.55/hr500 ms$0.0000764
NVIDIA L4 24GB$0.40/hr700 ms$0.0000778
NVIDIA T4 16GB$0.20/hr1500 ms$0.0000833

The per-image GPU cost converges around $0.00008 regardless of which card you pick, because the slower cards are cheaper per hour. The interesting number is not per-image but per-month.

Realistic monthly bill for a self-hosted LaMa shop processing 100,000 images:

If your usage is bursty (10,000 images on Monday, nothing Tuesday through Sunday), you can use a serverless GPU product like Replicate or Modal, which bills per second. At Replicate’s published LaMa price of $0.0007 per image (serverless cold-start included), 100,000 images is $70/mo of pure compute, plus whatever orchestration you build around it.

Putting the three together

The break-even is the question that matters. The browser tool is free; SaaS and self-host both have fixed costs. We compare per-image cost across realistic volumes:

Volume (images/mo)Browser toolSaaS Pro $9.99Self-hosted A10G 24/7Replicate LaMa
100$0.00000001$0.10$6.21 (over 100x)$0.07
1,000$0.0000001$0.010$0.62 (over 100x)$0.0007
10,000$0.000001$0.001$0.062$0.0007
100,000$0.00001$0.0001$0.0062$0.0007
1,000,000$0.0001$0.00001 (at cap)$0.000621$0.0007

The browser column is mostly a joke at these volumes — it is literally less than the rounding error of any other line. SaaS at the Pro tier crosses below $0.001 per image only at 10k images per month or more. Self-hosted wins at scale because the GPU is paid for whether you use it or not, but you have to actually use it 24/7 to get the per-image number above.

The interesting number is the crossover between SaaS and self-hosted. SaaS at $9.99/mo for 2,000 images is $0.005 per image. Self-hosted at $621/mo for 100,000 images is $0.0062 per image. The crossover volume is roughly 50,000 images per month. Below that, SaaS. Above that, self-hosted.

The hidden costs nobody puts on the page

Every option has costs that the price page does not show:

What the right answer looks like

There is no single right answer. Three honest recommendations:

FAQ

Why is the browser tool not $0 exactly?

It is $0 to the operator of the tool and effectively $0 to the user. The electricity cost of running the inpainter on a phone is less than one ten-thousandth of a cent per image, which is below the resolution of any monthly bill. We use the rounded figure $0.0000001 to make the comparison honest, not to suggest you will ever see this on a bill.

Do I pay for bandwidth in the browser tool?

Yes, but it is fixed, not per-image. The browser tool’s HTML, CSS and JS are a few hundred KB. They load once when you visit the page. After that, every image you process uses zero additional bandwidth, because the inpainting runs on your device. A SaaS tool charges per-image because every image is an upload to their server.

What about free tiers?

Free tiers usually mean: 3-5 images per month, watermarked output, or rate-limited to 1 image per minute. They exist to convert you to a paid tier, not to be a sustainable way to process images. A free tier that is actually free forever (like this site) is paid for by ads, by donations, or by the operator’s day job. Read the small print to find out which.

Does the GPU choice really not matter per image?

It matters less than people think. A faster GPU costs more per hour but finishes each image faster; a slower GPU costs less per hour but takes longer. For LaMa at the typical 200-1500 ms range, the per-image cost is dominated by the “GPU idle cost amortised over actual work” equation, and that converges to $0.00008-ish for any reasonable card. The choice between A100 and T4 is mostly about which workloads you want to handle, not about per-image cost.

What is the cheapest legal way to remove a watermark?

If you have the right to edit it, a browser tool is the cheapest, with the privacy bonus. If you do not have the right to edit it, the cheapest answer is still no, because infringement damages are not in the cost model above.

Methodology and sources

Electricity: US EIA Average Retail Price of Electricity, residential, July 2026 release. SaaS pricing: observed across the segment in 2026, anonymised. GPU spot pricing: AWS, GCP, Lambda Labs and RunPod published rates for us-east-1 equivalents in Q3 2026. Runtimes: 200 ms for LaMa on an A100 from the official LaMa repo benchmark, 500-1500 ms on A10G/L4/T4 measured on the Replicate hosted LaMa model. Operator time estimate for self-hosted is a midpoint between 1 hr/mo (well-run, autoscaled, no surprises) and 8 hr/mo (managed badly with weekly incidents).