|Docs

Enhance doorbell or dashcam footage: why invideo says no

invideo does not enhance doorbell, CCTV, security camera or dashcam footage to read a license plate, a number plate or a face. Its upscalers make a video look sharper by inventing detail, so a plate or a face they draw is a guess, not what the camera recorded. What invideo does is help you send the footage you have: a short trimmed clip, clear stills of the moments that matter, and a contact sheet, ready for the police or your insurer.

Use invideo for this when you need a short clip and clear stills to send with a police or insurance report, and you want to keep the original file untouched.

Quickest path

Set up your agent first with invideo for AI agents. Then give it this prompt.

I have doorbell footage of a package theft in ./doorbell.mp4. Do notenhance or upscale it. Trim it to the theft, give me stills at theclearest moments, and a contact sheet of the whole clip, so I can sendthem to the police with the original file.

In a chat app, you drag the file into the project's Media panel in invideo, then name it in the chat. In a terminal, you name the file and the agent uploads it.

What you need

Keep the original file exactly as the camera or the app saved it.

  • The original file. Download it from the camera app at full quality. Keep a copy that nobody edits.
  • The date, the time and the place. Your report needs them. The camera app usually shows them.
  • The moments that matter. A rough time in the clip for each one.
  • The report's address. The police report number, or your insurer's claim number and upload page.

Steps

These are the steps your agent does. It does not run an upscale or an enhance step on this footage.

  1. Bring in a copy

    The agent makes a project with new_project and uploads a copy of the file with start_upload and add_media. You keep the original. Expected result: a sequence URL for the copy.

  2. Read the file's details

    The agent runs media metadata for the file's length, size, frame rate and streams. Expected result: the details, which you can add to your report.

  3. Look through the clip

    The agent runs media video.frames at times across the clip, and video.contactSheet for one image of frames in a grid. It runs audio.content if the clip has sound, to find voices or a crash. Expected result: the times of the moments that matter.

  4. Pull clear stills

    For each moment, the agent runs media video.extractFrameAtTime with time_seconds. Each still is the frame as recorded, with nothing added. It sends an idempotency_key with each one. Expected result: one image sequence per moment.

  5. Trim a short clip

    The agent runs media video.trim with start_time and end_time around the event, a few seconds before and after, and an idempotency_key. It does not stabilize, sharpen, upscale or color the clip. Expected result: a short video sequence.

  6. Hand it over

    The agent gives you the URLs of the trimmed clip, the stills and the contact sheet. You download them from the project and send them with the original file, which stays the evidence. Expected result: files you attach to the report.

The same job from a terminal:

invideo project new --title "Doorbell 12 May"invideo agent register <project-url> --name <name>export INVIDEO_AS=<name>invideo asset upload <project-url> ./doorbell.mp4invideo media metadata <sequence-url>invideo media video.frames <sequence-url> --at 10,14,18,22 --out ./framesinvideo media video.extractFrameAtTime <sequence-url> --time-seconds 14.2 \  --dry-runinvideo media video.extractFrameAtTime <sequence-url> --time-seconds 14.2 \  --max-credits <price> --idempotency-key doorbell-still-1invideo media video.trim <sequence-url> --start-time 8 --end-time 26 \  --dry-runinvideo media video.trim <sequence-url> --start-time 8 --end-time 26 \  --max-credits <price> --idempotency-key doorbell-trim-1invideo status <still-sequence-url> <trim-sequence-url> --waitinvideo notebook show <still-sequence-url> --out ./reportinvideo notebook show <trim-sequence-url> --out ./report

Where each value comes from:

  • <name> is the name you choose in invideo agent register. INVIDEO_AS passes it to every later command.
  • <project-url> comes from invideo project new. <sequence-url> is the URL invideo asset upload returns for the copy. The stills and the trim make files, so they take that URL, not a media bin id. <price> is the credits value of each dry run.
  • <still-sequence-url> is the URL video.extractFrameAtTime returns, and <trim-sequence-url> the URL video.trim returns.
  • invideo status waits and saves nothing. invideo notebook show with --out saves each file to the folder.

The key MCP call is the full media call. It pulls one still at the time the face or the plate is clearest, as recorded. <sequence-url> is the URL add_media returns for the copy:

{  "url": "<sequence-url>",  "op": "video.extractFrameAtTime",  "time_seconds": 14.2,  "idempotency_key": "doorbell-still-1"}

The contact sheet uses video.contactSheet on the same sequence URL. It needs all four of columns, rows, resolution and style, and style is labeled or plain. A dry_run shows its price before it starts.

Cost and time

Edits and exports are free. Generation and transcripts cost credits. Paid stock costs a small credit charge, or one of your workspace's iStock licenses. Your agent runs a dry run first and tells you the price before it spends anything.

  • Free: the upload and the download.
  • Tool ops: metadata, video.frames, video.contactSheet, audio.content, video.extractFrameAtTime and video.trim. The dry run shows its price before anything starts.
  • Costs credits: a transcript of speech in the clip with audio.transcript, and the face masks with sam_3_1_video for the last prompt below.

The whole job takes a few minutes. See Credits and what they cost for prices.

If something goes wrong

Most questions on this job are requests invideo turns down, and why.

SymptomFix
get_model or generate refuses the model as not foundModels retire. Run find_models with the job in words, use the best match, and tell the person which model you used.
Exit code 6, or status is heldA safety review, not a refusal. Wait, then run invideo status <url> --wait again. Do not resend. If it ends failed, change the input the review flagged.
You ask to read a plate or a face that is a blurThe agent says no. An upscale invents the detail, and a made-up plate can name the wrong person. Ask the police or your insurer.
You ask who the person isThe agent does not identify people from footage. Send the stills to the police.
You ask to upscale "a little" for the reportThe agent says no for evidence. It sends the frames as recorded.
The camera app saves a low quality copyDownload the original again at full quality from the camera app or its cloud.
A still is blurred by motionThe agent pulls the frames before and after with video.frames, and you pick the clearest.
CLI exit 7Nothing ran. Run the same command again.
Avoid: sending only the trimmed clipInstead: send the original file too. Investigators need it as the camera saved it.
Avoid: posting an "enhanced" still onlineInstead: post the frame as recorded, or none. A wrong guess can harm someone.

Prompts to try

Each prompt is a real version of the job.

My dashcam caught a car hitting mine in ./dashcam.mp4 at about 2:10.Trim from 2:00 to 2:30, pull stills every second of the impact, and acontact sheet. No enhancement.
Make a contact sheet of my CCTV clip ./yard.mp4 and tell me the timeswhen someone is in the frame. Do not upscale anything.
I want to share the doorbell clip in my neighborhood group. Trim it,and blur the faces of my neighbors who walk past, but not theperson at the door.

For the last prompt, the agent masks the neighbors with sam_3_1_video or object_track, then blurs them with track_matte_key and gaussian_blur, priced first. It does not sharpen the person at the door.

When to use something else

Forensic video enhancement and identification are jobs for the police and for the forensic lab your insurer uses. They work from the original file, record every step, and their findings stand up in a claim or in court. Give them the original file and your report number.

If the camera company holds a cloud copy, ask them for the original export too. It can have more detail than the copy on your phone.