Comparison · Open source vs API

Subtitle Removal API vs Self-Hosted Inpainting

Open-source inpainting models exist, and they are genuinely capable. The gap between 'the model runs' and 'the service runs reliably at scale' is where the API wins — and where most self-hosting cost actually hides.

InpaintingSelf-hostingComparison

What open source gives you

Models like E2FGVI, ProPainter and STTN are published with weights and papers, and community forks run them on consumer GPUs. For a fixed, controlled workload, self-hosting is absolutely viable — many teams do it well.

The real cost breakdown

Self-hostedmlslabs API
GPU hardware1× A100/4090+ per lane; idle cost between bursts$0 — we scale workers to your queue
VRAM per job6–16 GB depending on model & resolutionNone on your side
Model opsCheckpoints, versions, retraining glueHandled; quality shipped continuously
Queue & retriesYou build themNative async jobs, webhooks, auto-retry
Cost shapeFixed monthly infra + your timePer minute of video, failed jobs refunded
Data controlFull — media never leaves your envMedia leaves your env (S3 in/out)

When self-hosting is right

Data-residency mandates, air-gapped pipelines, or an existing GPU fleet with spare capacity. If the compute is already paid for and idle, marginal cost approaches zero.

When the API is right

Variable volume, tight deadlines, teams without GPU ops, or workloads where a queue, retries and webhooks matter more than the model itself. The API also removes the 'who retrains and when' question from your roadmap.

Honest numbers

Note: We will not publish fake benchmarks. Run your own test: our free tier covers a pilot batch; measure wall-clock time, quality on YOUR footage, and the full monthly cost including the hours your team spends on GPU maintenance. That number is the comparison that matters.

FAQ

Common questions

Can I run inpainting on my own GPU?

Yes — open-source video inpainting models run on consumer GPUs. Expect to own the whole service layer: queues, retries, monitoring, model updates.

Is the API cheaper than a GPU machine?

At low volume, a free-tier API is cheaper; at very high sustained volume, owned GPUs can win. The break-even depends on utilization — the API's advantage is paying only for what you use.

What if my media must stay on-prem?

Then self-host, or talk to us about private deployment. The API defaults to hosted processing with object-storage transfer.