Documentation/Practical guides

Practical guides

Cost of AI image and video watermarking on CPU

Ariadne Mark has no software licence fee or per-export Ariadne API charge. You run the toolkit on your infrastructure. Your operating budget covers compute, hosting, storage, verification and any production signing or timestamp services you use.

Measured CPU work for 1,000 daily exports

The following illustration uses the 0.7.3a4 measurements from 15 November 2026: a 1024 × 683 PNG and a ten-second 720p/30 fps engineering video. The EC2 m6a.large container was limited to two CPUs and 4 GiB. Warm values are medians of three resident runs after one first job. CPU includes Node, native codec work and completed child processes.

Measured input / profile Warm export, wall time Warm total CPU, median CPU-hours for 1,000 identical exports
1024 × 683 PNG 0.61 s 0.98 CPU s 0.27 CPU h
10 s 720p · Economy5s 9.16 s 13.16 CPU s 3.66 CPU h
10 s 720p · Economy2s 13.85 s 21.49 CPU s 5.97 CPU h
10 s 720p · Resilient 107.64 s 189.35 CPU s 52.60 CPU h

The last column is arithmetic: 1,000 × measured CPU seconds / 3,600. It is CPU-hours, not elapsed hours, a throughput promise or a cloud invoice. Different media, concurrent jobs and production signing can change the result. Read the full protocol and raw data.

Estimate your own worker budget

Use:

CPU-hours/day = exports/day × measured CPU seconds/export ÷ 3,600

For actual capacity, measure completed jobs per hour at your chosen worker/concurrency limit. Leave room for spikes, retries and verification. Two vCPUs do not guarantee twice the throughput, and an always-on VM is usually billed for its allocated time rather than only busy CPU seconds.

Separate the first job from resident repeats. The Python integration reuses an SDK inside a worker; a per-file CLI subprocess may pay more startup overhead. A Node.js app can keep the same queue while delegating media work to that worker.

What else costs money?

Item What to provision or measure
Export workers Cold/warm wall time, CPU and memory for your actual dimensions, duration, FPS and codec
Verifier Capacity for independent checks, long negative scans, concurrency and bounded uploads
Storage and registry Published media, receipts, manifests, recovery records and backups
Signing identity Eligibility, certificate renewal, key handling and any conformance work
Trusted timestamps Required request volume, vendor quota or price, latency and retry behavior
Delivery CDN/app transformations and any separate final renditions that must be marked

The public image playground is a demo, not a production marking service or a hosted verifier SLA for your app. Deploy and size your own endpoint through the connected kit.

A complete workload: 1,000 images and 300 videos per day

At one final rendition and one successful export per item, a 30-day month is 30,000 image exports + 9,000 video exports = 39,000 exports. A 365-day year is 474,500 exports. Download the editable workload sheet; unknown inputs are null, never silently zero. These are planning counts, not a claim that one VM can process this workload.

export attempts = days × (images/day + videos/day) × renditions/item × attempts/rendition
export CPU-hours = (image attempts × full-job image CPU seconds
                 + video attempts × full-job video CPU seconds) / 3600
verify CPU-hours = days × (positive checks/day × full-job positive CPU seconds
                 + negative checks/day × full-job negative CPU seconds) / 3600
monthly compute bill = exporter allocated VM-hours × its hourly price
                     + verifier allocated VM-hours × its hourly price
monthly total = compute + storage/backup + delivery + domain/TLS
              + certificate + TSA + operator work + allocated setup costs

Measure CPU for the whole job/process tree, including Node and FFmpeg. VM billing uses allocated hours; CPU-hours above describe work and cannot be directly multiplied by a VM-hour price. Capacity still needs a sequential/queued wall-time measurement with headroom. Keep image and each selected video profile separate. The recent client pilot's Python SELF CPU excludes child processes; it is not a valid complete CPU input or a basis for a dollar total.

Budget line Fill in for this workload Not included automatically
Export Image/video dimensions, duration, FPS and profile; cold/warm wall time, total CPU and peak memory; retries/renditions Generation GPU/API cost, throughput guarantees
Positive verification Checks/day × measured positive job cost Export gate is not your public independent verification traffic
Negative/incomplete verification Clean negatives/day × full-scan cost; budget/deadline and queue headroom Early positive timings cannot size this load
Hosting Allocated exporter and separate verifier hours × actual regional VM prices Idle time, scaling and network egress are not free
Storage/registry/backup Media bytes × retention; receipt/manifest/database growth; tested backup/restore A deleted upload is not a backup policy
CDN/delivery Egress and each final crop/encode's additional export Original master export does not mark a new rendition
Signing/TSA Certificate term + requests above real quota + signing retries Failed export may still consume a timestamp; exact retry volume needs measurement
Operations/setup Engineer hours × your rate; enrollment, renewal, monitoring, incidents Free SDK code does not supply managed operation

Do not add the bundled timestamp service again if already included in a managed plan. Separate recurring spend from initial implementation/conformance work. Taxes, currency conversion, discounts and temporary promotions are excluded here.

Dated external price references

Checked 15 November 2026; confirm eligibility, interface and current terms before ordering. These references do not establish interoperability with your credentials, a legal approval or an all-in quote.

For a hosting reference, AWS Lightsail lists a Linux public-IPv4 bundle at $12/month with 2 GB RAM, two burstable vCPUs, 60 GB SSD and a 3 TB transfer allowance (regional transfer conditions apply). This is a small site/demo hosting reference, not a qualified worker size or capacity estimate for 39,000 image/video exports. Do not count both a bundled hosting fee and the same VM's hours twice.

SSL.com's C2PA pricing lists a one-year Level 1 certificate at no charge with 2,500 timestamps/year; its Premium lists $500 per certificate/year with 10,000 timestamps/year. At 474,500 annual successful exports, those allowances leave respectively 472,000 or 464,500 requests beyond quota before retries. Overage needs a quote. Eligibility/conformance and key-interface compatibility remain operator work.

Trufo's public pricing lists Managed Starter at $120/month including 1,000 signatures, then $0.03/signature; Business at $1,200 including 10,000, then $0.01. Applied mechanically to 39,000 signatures, these give $1,260 and $1,490/month for the signature-plan component. They are not full workload quotes: included watermark ID units are 300/3,000; video unit accounting and overage need confirmation. Volume discounts or a temporary promotion can change the result. Managed service includes operator work that self-hosted Ariadne leaves to your team; prices alone do not establish which option is cheaper for your app.

Before choosing a budget

Fill every unknown line, test representative files on your intended worker and confirm a CA/TSA quote for the request volume. A partially filled worksheet can show workload and known components, but cannot report an all-in monthly total. The production critical path identifies which steps need your own operator credentials and an external service.

Which video profile uses less compute?

Economy5s and Economy2s reduce inference to selected marking windows. Resilient computes a fresh signal on every frame and used more CPU in this test. All three still decode and encode the complete file.

Choose with both cost and delivery behavior in view. Short excerpts can miss Economy windows; cropping and removal can defeat recovery in any profile. The mode selector, profile documentation and evidence describe those limits. A low-cost mode is not a separate legal certification.

Production signing is a separate setup step

The SDK's production export policy currently requires trusted timestamps. This is a toolkit implementation choice, not a statement that every law mandates this exact per-file TSA architecture. Obtain terms and capacity from the certificate/TSA provider you actually use; the dated price components above are references, not an unlimited free quota or a deployment quote.

The real identity/enrollment and authenticated TSA route still need qualification for your deployment. Start with production onboarding and account for external service costs before committing to a launch date.

The current SDK supports bearer-authenticated timestamping through its bounded loopback relay. Its offline fixture tests do not enroll your company or qualify live service quotas and latency. Supply actual operator credentials and verify a real production export before treating timestamp cost as the only remaining work. See the integration checklist.

Compare operating models · Measure a ten-second video · Run without a GPU · Get started