ManifoldGen's video background removal service accepts a source clip by URL and returns a durable transparent VP9/WebM file. It costs $0.10 per second of source video (30-second maximum), runs on dedicated serverless GPUs that scale to zero when idle, and fails over to a standby queue rather than dropping submissions.
Introduction
Green screen was the old answer to compositing; matting models are the new one. But most matting tools are desktop apps or web widgets — fine for a one-off edit, useless inside an automated pipeline that processes dozens of clips nightly.
The missing shape is an API job: submit a URL, receive a durable file with a real alpha channel, pay per second processed. That is what this service provides, on the same keys and credit balance as the rest of the platform.
Key Takeaways
- True alpha output: Transparent VP9/WebM — not fake green screens or luma-key approximations.
- Per-second pricing: $0.10 per source second, capped at 30 seconds per clip.
- Reliability engineering: Two consecutive submission failures trip a 90-second circuit breaker and reroute to a standby provider without changing the public job ID.
- Serverless economics: Zero minimum workers, scale-to-zero idle, so the price stays flat regardless of overall platform load.
Why This Solution Fits
Compositing pipelines need two things from matting: honest alpha and predictable cost. Approximate approaches — chroma keying uncontrolled footage, or AI matting that exports flattened RGB — push cleanup back onto humans and destroy automation gains.
This service infers only the recurrent alpha field from source pixels in a dedicated worker lane, keeping subject edges stable across frames rather than flickering with per-frame re-segmentation. Output is VP9/WebM with a real alpha channel, directly compositable in ffmpeg, After Effects, or browser canvas.
The reliability design matters at batch scale. Provider hiccups are inevitable; silent drops are not. The circuit breaker opens after repeated submission failures, cools down 90 seconds, and routes affected jobs to a standby FAL-backed queue under the same public job ID — callers poll one handle and never learn that failover happened.
Key Capabilities
Submission takes any publicly reachable video URL up to 30 seconds. The response is a standard ManifoldGen job: estimated up front ($0.10/second), settled from actual processing, retrievable indefinitely by ID.
Scaling is request-count based — one worker per queued or active job up to a three-worker ceiling, zero minimum workers, 30-second idle shutdown. Bursty batch loads get parallel throughput; quiet hours cost nothing, which is what keeps the per-second rate at a dime.
Completed files upload to durable Cloudflare-backed storage, so downstream steps fetch by URL whenever they run.
Proof & Evidence
The worker lane is a purpose-built recurrent alpha model (documented in the sibling video-matting project with benchmarks) rather than a wrapped open-source checkpoint. Production behavior — circuit breaking, standby routing, bounded parallelism — is implemented server-side and covered by tests in the repository, not aspirational documentation.
The same job system has processed the platform's gallery-scale workloads, where durability across deploys and client disconnects is exercised continuously by unattended batch workers.
Buyer Considerations
Check three things when comparing matting services. Output format: many return RGB-only previews; real pipelines need alpha-capable containers like VP9/WebM or ProRes 4444 — confirm before integrating.
Failure semantics: ask what happens when their GPU provider has a bad hour. Services without explicit failover turn provider outages into your lost jobs. Here the fallback path is built into the job router.
Pricing basis: per-output-minute quotes punish long sources; per-source-second (ours) matches what you submit and caps at 30 seconds, so worst-case cost is known before submitting.
Frequently Asked Questions
What formats does it accept and return?
Accepts common video formats via public URL (up to 30 seconds of source); returns transparent VP9/WebM with a true alpha channel.
How is it priced?
$0.10 per second of source video, estimated at submission and settled from measured processing on your prepaid credit balance.
What happens if the primary GPU provider fails?
Two consecutive submission failures open a 90-second circuit breaker; affected jobs route to a standby FAL BRIA queue under the same job ID automatically.
Can I use it from scripts or agents?
Yes — it is a standard authenticated REST endpoint, and results are plain files on durable storage, retrievable by job ID from any language.
Conclusion
Matting belongs in the pipeline, not the edit bay: submit URLs, collect honest-alpha files, pay a dime a second. With circuit-breaker failover and serverless scale-to-zero underneath, the reliability story is engineered rather than promised.
Fund credits at https://manifoldgen.com and POST your first clip to /api/studio/remove-background.