The Synthetic Video Detector (SVD) flags AI-generated or manipulated video on a live stream, backed by NVIDIA's Maxine Synthetic Video Detector model. Set a stream's detector type to Synthetic, and VIF returns a per-window real/fake verdict against a configurable threshold.
It's opt-in and bring-your-own-endpoint; nothing runs unless a stream is configured for it and pointed at a reachable SVD endpoint, either a local detector sidecar VIF can bring up for you, or any hosted/self-hosted endpoint you already run.
Licensing: The model is distributed through NVIDIA NGC / NVIDIA AI Enterprise and is access-gated, which means you authenticate with your own NVIDIA key. Wowza doesn't redistribute the model image or weights.
Things to know before deploying
- Video quality is untouched unless you opt into the verdict overlay, which adds a separate transcoded rendition rather than altering your source.
- Verdicts trail live by about one analysis window, since a window must fully capture before it can be analyzed.
- The GPU requirement is on the detector host only, not your streaming engine. SVD needs NVENC/NVDEC hardware plus Tensor cores, which excludes datacenter training GPUs like the A100, H100/H200, and B100/B200. Consumer/prosumer cards with NVENC/NVDEC (T4, A10, L4, RTX 4090/5090, and similar) generally work. Check the repo's GPU support matrix before provisioning hardware.
- The source must be H.264. A non-H.264 source errors clearly rather than being transcoded.
- If the detector falls behind, VIF skips forward rather than queuing. It always analyzes the most recent completed window, so latency stays bounded, but coverage can have gaps. A per-window timeout (default 60s) prevents one hung request from blocking the stream.
Deployment
There are two deployment paths. A local detector sidecar brought up as part of the Docker Compose stack, or pointing a stream at a hosted or self-hosted endpoint you already run (including NVIDIA's hosted dev endpoint, which is evaluation-only, not production). The sidecar and the service querying it don't need to share a machine.
Securing a remote deployment: A self-hosted detector doesn't check API keys, and only NVIDIA's hosted endpoint does. If you publish the sidecar's port for remote access, mutual TLS is the only built-in way to control who can reach it; see the repo's TLS/mTLS setup before exposing a detector beyond a single machine.
What you receive
Each analysis window produces one result: a verdict ("synthetic", "real", or "unknown" if the endpoint was unreachable or rejected the request) and a synthetic_score between 0.0–1.0. Results are delivered through the same event listeners as every other detector type: log file, ID3 tags, and webhooks by default.
The verdict overlay is off by default — it requires transcoding, which the default configuration doesn't do. Enable use_transcoder on the stream if you want the overlay rendition too.
Threshold: the default (0.3) is conservative, minimizing missed synthetics at the cost of some false positives. Use 0.5 for a more balanced trade-off, and tune to your own review policy.
Compliance note (EU AI Act Article 50)
Article 50 of the EU AI Act requires transparency obligations for AI-generated or manipulated content, applicable as of 2026-08-02. The detector is a decision-support signal, not a legal determination; tune the threshold and your review policy to your own obligations.
Air-gapped deployment
SVD runs fully offline at inference time. The only network step is a one-time model download on first boot. See the repo's air-gapped section for pre-seeding the model cache on a disconnected host.
