ComfyUI-VideoGate: The QC Gate That Stops Broken AI Video Before It Ships

The most expensive bug in AI video isn't in the model. It's in the delivery.

You render overnight. The queue says "done." You send the folder to the client. And then the client replies: the last clip is ten seconds of black. The vertical one isn't vertical. And where's the audio?

Every AI video creator has a version of this story. The render pipeline has every incentive to say "finished" — and nothing in the pipeline is paid to ask "deliverable?"

I've been building that question as the AIVideo safety net for a while: QualityGate to gate, AdherenceGate to check briefs, CreditGuard to watch spend, BatchQueue to survive the night. But they all lived outside the tool where AI video actually gets made: ComfyUI.

So I ported the gate to where the renders happen. ComfyUI has ~61,000 custom nodes across 8,000 extensions — image QC gates, deflicker fixers, VLM "describe my video" tools. What it didn't have: a plain, honest, offline gate that reads a finished render and says PASS or FAIL before anything leaves the machine.

ComfyUI-VideoGate

Four nodes, category video/Videogate, zero GPU, zero API keys — just ffprobe and ffmpeg, which every render box already has:

• VideoGate Probe — every stream fact: container, codecs, WxH, fps, duration, frame count, aspect, audio presence
• VideoGate Check — gate against a delivery profile (client_delivery, social_vertical, archival_master…) plus black-frame, frozen-frame, and EBU R128 loudness checks → a verdict and machine-readable issues JSON
• VideoGate Report — JSON / Markdown / CSV QC reports next to the clip
• VideoGate Enforce — the branch: forward the video on PASS/WARN, or hard-stop the workflow so a broken clip physically cannot continue

LoadVideo ─▶ VideoGate Check ─▶ VideoGate Enforce ─▶ (deliver)
                  │
                  └▶ VideoGate Report ─▶ report.json / report.md

The typical failure modes of AI video render — the black tail from a truncated run, the frozen frames where generation stalled, the 16:9 export that was supposed to be 9:16, the silent "with music" clip — all get caught in-graph, in seconds, offline.

What building it taught me about ffmpeg

The pack is small, but ffmpeg fought back in ways worth recording:

1. blackdetect changed its field names. ffmpeg ≥ 7 emits black_duration; older builds emit black_dur. Parse only one and you silently report "no black frames" on a clip that's half black — the worst kind of bug: a QC tool that passes broken video.
2. freezedetect's noise parameter is now a 0–1 ratio, not dB. Passing the old -60 crashes the filter outright on current builds.
3. freezedetect only reports a freeze's duration when motion resumes. A freeze that runs to end-of-file never gets its end marker — an end-of-clip stall is invisible to the detector. My test fixture had to put the freeze mid-stream (animated → frozen → animated) to be detectable at all. Real-world lesson: always chase the tail of the render; that's exactly where generation stalls.
4. freezedetect prints a bare freeze_start: marker line before a confirmed event — a naive parser counts the marker itself as a freeze, flagging clean clips. The gate now requires a real duration on every event.

Every one of these was found by testing against real generated media — testsrc2 patterns, tpad-cloned freezes, concat splices — never mocks. The suite (33 tests) builds its own fixtures with ffmpeg at test time and runs the actual node functions. If the ffmpeg on the box changes behavior, the tests catch it, not the client.

Where it fits

VideoGate is the sixth pillar of the safety net, and the first one that lives inside the render tool:

• Gate inside ComfyUI → ComfyUI-VideoGate (this)
• Gate from the terminal → python -m videogate.cli gate clip.mp4
• The original CLI gate → AIVideoQualityGate
• Automations on the verdict → the Check node's issues_json output feeds webhooks / n8n / xyOps directly

The node pack accepts ComfyUI's native VIDEO type and plain path strings, so it wires into LoadVideo/SaveVideo pipelines and into any custom node that emits file paths.

Try it

git clone https://github.com/madebysaira/ComfyUI-VideoGate.git
# drop into ComfyUI/custom_nodes/  (needs ffmpeg on PATH)

Or once it's on the Comfy Registry: comfy node install comfyui-videogate

A render that finishes is not a render you can deliver. Now there's a node for the difference.