Check a video

How a video runsCopy Anchor Link

A video check is a job. POST /audit answers 202 at once with its id and media: "video", and you poll GET /audit/{id} until its status is succeeded or failed.

The job moves from queued to running, then ends as succeeded or failed. Poll it every few seconds. Each poll counts toward your requests a minute, so on 429 rate_limited wait the seconds in Retry-After, then poll again.

Send exactly one of video_url or video_upload_id. A video on disk goes through an upload with its video content_type. A large video is faster that way than by URL.

Send the lengthCopy Anchor Link

Pass seconds, the video's length, when you know it. Credits are held for that length and a video that runs longer is refused. Without it, credits for the longest video are held, and the real length is charged once the job finishes. Video limits has the numbers.

Check a videoCopy Anchor Link

curl -X POST "https://realpng.com/api/v1/audit" \
  -H "X-API-Key: $REALPNG_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "video_url": "https://videos.example.com/drone.mp4",
    "seconds": 12
  }'

{
  "id": "b8f2d6a4-3c1e-4f7a-9d5b-2e6c8a0f4b13",
  "status": "queued",
  "media": "video"
}

Then poll the check until it finishes.

curl "https://realpng.com/api/v1/audit/b8f2d6a4-3c1e-4f7a-9d5b-2e6c8a0f4b13" \
  -H "X-API-Key: $REALPNG_API_KEY"

{
  "id": "b8f2d6a4-3c1e-4f7a-9d5b-2e6c8a0f4b13",
  "status": "succeeded",
  "verdict": "ai",
  "score": 88,
  "decided_by": "pixels",
  "lanes": {
    "pixels": {
      "verdict": "ai",
      "score": 88
    },
    "watermark": {
      "verdict": "none",
      "score": 4
    },
    "provenance": {
      "verdict": "nothing",
      "issuer": null
    }
  },
  "media": "video",
  "video": {
    "seconds": 12,
    "fps": 30,
    "frames": 360,
    "width": 1280,
    "height": 720,
    "audio": "copied",
    "readings": [
      {
        "t": 0.75,
        "verdict": "ai",
        "score": 84
      },
      {
        "t": 2.25,
        "verdict": "ai",
        "score": 86
      },
      {
        "t": 3.75,
        "verdict": "ai",
        "score": 88
      },
      {
        "t": 5.25,
        "verdict": "ai",
        "score": 91
      },
      {
        "t": 6.75,
        "verdict": "ai",
        "score": 87
      },
      {
        "t": 8.25,
        "verdict": "ai",
        "score": 90
      },
      {
        "t": 9.75,
        "verdict": "ai",
        "score": 89
      },
      {
        "t": 11.25,
        "verdict": "ai",
        "score": 85
      }
    ]
  }
}

A video is read on 8 frames, one at the middle of each of 8 equal slices of the video. Each frame is checked as an image is, both detectors and the watermark check, and the video's reading is the median of the frames'. One scene cut or black frame does not decide it.

verdict, score, decided_by and lanes mean what they mean for an image, as Read the verdict describes. The provenance lane reads the video file's own C2PA manifest. video adds the file's facts and each frame's reading.

The frames of one video can read far apart, so branch on the video's verdict, not on one frame's.

From an agentCopy Anchor Link

The MCP tool audit_video takes the same fields and waits for the job, as MCP tools describes. get_processing_result resumes it.

Next stepsCopy Anchor Link