Platforms aren't downranking your AI clip because it looks bad. They're downranking it because it looks obvious. Three signals trigger the AI label in 2026: visible watermarks burned into the frame, C2PA Content Credentials riding in the file header, and SynthID's invisible pixel-level watermark embedded by Google's video models. Strip all three and the same clip stops getting buried.
This post is the working approach as of mid-2026. Each signal needs a different removal step; doing one without the others is the most common reason creators say "I scrubbed the watermark and it's still getting flagged."
1. Visible watermarks
Some video engines (older Runway, Pika free tier, a few Wavespeed presets) burn a small logo into a corner of the last frame or across the full duration. Platform classifiers match the logo shape itself, so even a blurred-over version gets caught.
The fix is region-detection plus inpainting across every frame. The watermark sits in the same pixel coordinates throughout the clip; isolate the bounding box, mask it, and run a content-aware fill pass that interpolates from the surrounding pixels of the previous and next frames. FFmpeg's delogo filter is a fast first pass; a deep-learning inpainter (LaMa, ProPainter) handles the tricky cases where the watermark sits over a face or moving object.
2. C2PA Content Credentials
C2PA is a metadata standard pushed by Adobe, Microsoft, and OpenAI. It rides in the file header as a signed manifest declaring which AI model produced the asset and when. Instagram and YouTube both read it on upload; if present, the post is auto-labeled "AI-generated" with no override path.
Removal is a single ffmpeg pass that copies the streams without the metadata blocks:
The -map_metadata -1 flag drops every metadata track, including the C2PA manifest. -c copypreserves the video and audio streams without re-encoding, so there's no quality loss and the operation takes under a second.
3. SynthID
SynthID is Google's invisible watermark for Veo, Imagen, and Lyria output. It modulates pixel values in patterns that are imperceptible to humans but detectable by Google's own classifier — and increasingly by third-party detectors as the API spreads. Stripping metadata doesn't touch it because it lives in the pixels themselves.
The working removal as of mid-2026 is a targeted denoise-plus-grain pass. Apply a mild Gaussian noise to break the SynthID pattern, then re-apply a structured grain that matches camera-sensor noise. The signal is gone; the clip looks like footage from a phone in low light.
The 1% grain pass
After visible watermarks, C2PA, and SynthID are handled, one signal remains: the texture-flat look of synthetic video. AI clips have a characteristic absence of sensor noise that platform "look classifiers" use as a tell. A 1% film-grain pass — barely visible to the eye — adds the kind of noise a real camera produces.
FFmpeg recipe:
alls=8 is the noise strength (8/100 = ~1% in perceptual terms). allf=t makes the noise temporal — different every frame, the way real grain behaves.
Do it in this order
- Strip C2PA first (cheap, no quality loss).
- Remove visible watermarks (slow if you need the deep inpainter, but mandatory if any are present).
- SynthID denoise pass (only on clips from Google models).
- 1% grain pass as the final step before delivery.
What doesn't work
- Adding a sticker over the watermark. Mobile editors detect the original underneath; sticker just becomes a second signal.
- Re-encoding with a different codec. Neither C2PA metadata nor SynthID survive every re-encode, but the visible watermark does.
- Manually trimming the first/last second. The watermark is present every frame; trimming changes nothing.
- Hoping the audience doesn't notice. The platform classifier sees what humans miss — that's the entire job description.
Is this legal?
Stripping metadata from a video you generated yourself is legal in every jurisdiction we operate in. C2PA is an industry standard, not a law; SynthID is a Google policy signal, not a copyright claim. You own the videos you generate with ViralGen — and removing identification signals from your own assets is no different from filing the serial number off your own laptop.
Don't strip metadata from a clip you didn't generate. That crosses into removing attribution from someone else's work, which is a different conversation entirely.