DEV Community

RenderIO
RenderIO

Posted on • Originally published at renderio.dev

Make AI-Generated Video Undetectable on TikTok

#ffmpeg #video #api #webdev

How to make AI video undetectable on TikTok

You generated a video with Runway, Kling, Pika, or Sora. It looks great. You upload it to TikTok. It gets suppressed or flagged.

The problem is twofold: metadata fingerprints from the generation tool, and visual patterns that detection algorithms catch. Both are fixable with FFmpeg. This guide walks through each fix step by step, so you can make AI-generated video undetectable before uploading.

What makes AI video detectable

Metadata fingerprints

Every AI video tool embeds metadata in the output file. The encoder field contains the tool's name or rendering engine. Creation timestamps are typically UTC and batch-generated (a giveaway when you upload seconds after "recording"). Some tools add proprietary tags and custom metadata fields. C2PA Content Credentials — increasingly common in 2025-2026 — explicitly declare AI origin. And the EXIF data (resolution, color space, technical settings) often matches the tool's default output exactly, which is another signal.

For a full walkthrough on stripping video metadata with FFmpeg, that guide covers every metadata field and how to remove them.

You can see this with ffprobe:

ffprobe -v quiet -print_format json -show_format input.mp4
Enter fullscreen mode Exit fullscreen mode

Look for fields like encoder, comment, creation_time, and any tool-specific tags.

Visual patterns

AI video has characteristic patterns:

  • Consistent frame timing: AI renders at exact intervals. Natural video has micro-variations in frame timing.
  • Uniform noise patterns: AI-generated frames lack the random sensor noise present in camera footage.
  • Temporal consistency: AI maintains unnaturally smooth motion in areas that real cameras would show compression artifacts.
  • Color space: Many AI tools output in a specific color space (often BT.709 with particular gamma curves).

Step 1: Strip all metadata

Remove every metadata field:

ffmpeg -i ai_video.mp4 -map_metadata -1 -fflags +bitexact -c:v libx264 -crf 22 -c:a aac output.mp4
Enter fullscreen mode Exit fullscreen mode

-map_metadata -1 removes all metadata containers. -fflags +bitexact prevents FFmpeg from writing its own metadata.

API call:

curl -X POST https://renderio.dev/api/v1/run-ffmpeg-command \
  -H "Content-Type: application/json" \
  -H "X-API-KEY: your_api_key" \
  -d '{
    "ffmpeg_command": "-i {{in_video}} -map_metadata -1 -fflags +bitexact -c:v libx264 -crf 22 -c:a aac {{out_video}}",
    "input_files": { "in_video": "https://example.com/ai-video.mp4" },
    "output_files": { "out_video": "clean.mp4" }
  }'
Enter fullscreen mode Exit fullscreen mode

Step 2: Add natural sensor noise

Real cameras produce random noise from the image sensor. AI video is too clean. Add subtle noise:

ffmpeg -i ai_video.mp4 -vf "noise=alls=8:allf=t" -c:v libx264 -crf 22 output.mp4
Enter fullscreen mode Exit fullscreen mode

alls=8 adds noise at strength 8 across all color planes. allf=t makes it temporal (varies per frame), mimicking real sensor behavior.

For a more natural look, add gaussian noise instead of uniform:

ffmpeg -i ai_video.mp4 -vf "noise=alls=6:allf=t+u" -c:v libx264 -crf 22 output.mp4
Enter fullscreen mode Exit fullscreen mode

Step 3: Introduce frame timing variation

AI video has perfectly consistent frame timing. Real video from phones has slight jitter. Add micro-variations:

ffmpeg -i ai_video.mp4 -vf "setpts=PTS+random(0)*0.001" -c:v libx264 -crf 22 output.mp4
Enter fullscreen mode Exit fullscreen mode

This adds up to 1ms of random timing variation per frame. Invisible during playback but breaks the perfect timing pattern.

Step 4: Re-encode to match phone camera output

TikTok expects video from phones. Match the encoding characteristics:

ffmpeg -i ai_video.mp4 \
  -c:v libx264 -profile:v high -level:v 4.0 \
  -crf 23 -preset medium \
  -pix_fmt yuv420p \
  -c:a aac -b:a 128k -ar 44100 \
  -movflags +faststart \
  output.mp4
Enter fullscreen mode Exit fullscreen mode

This matches the H.264 High Profile Level 4.0 output that modern phones produce. yuv420p is the standard pixel format. movflags +faststart is how phone cameras write MP4 files.

Step 5: Crop to remove AI artifacts

AI videos often have subtle artifacts at frame edges (blurring, warping, or inconsistent generation). Crop a few pixels:

ffmpeg -i ai_video.mp4 -vf "crop=iw-8:ih-8:4:4" output.mp4
Enter fullscreen mode Exit fullscreen mode

This removes 4 pixels from each edge. Eliminates edge artifacts and changes the perceptual hash.

Step 6: Adjust color space

Runway and Sora output in BT.709 with a specific gamma curve (usually 2.2 or sRGB transfer). Kling defaults to BT.709 but with flatter gamma that gives a slightly washed-out look. Pika's output varies by model version. The point is: each tool has a default color profile that detection systems can fingerprint. Shift it:

ffmpeg -i ai_video.mp4 -vf "eq=brightness=0.02:contrast=1.02:saturation=1.03:gamma=1.01" output.mp4
Enter fullscreen mode Exit fullscreen mode

Slight brightness, contrast, saturation, and gamma adjustments. These shift the color profile away from the AI tool's default output.

The complete naturalizer command

Combine all steps into one FFmpeg command:

ffmpeg -i ai_video.mp4 \
  -vf "crop=iw-6:ih-6:3:3,noise=alls=6:allf=t,eq=brightness=0.015:saturation=1.02,hue=h=1" \
  -af "asetrate=44100*1.003,aresample=44100" \
  -c:v libx264 -profile:v high -level:v 4.0 -crf 23 -preset medium \
  -pix_fmt yuv420p \
  -c:a aac -b:a 128k -ar 44100 \
  -map_metadata -1 -fflags +bitexact \
  -movflags +faststart \
  output.mp4
Enter fullscreen mode Exit fullscreen mode

This command:

  1. Crops edges (removes AI artifacts, changes pHash)
  2. Adds sensor noise (naturalizes the image)
  3. Shifts brightness and color (moves away from AI defaults)
  4. Shifts audio pitch slightly (alters audio fingerprint)
  5. Encodes to phone-camera-like specs
  6. Strips all metadata
  7. Optimizes for mobile playback

API call:

curl -X POST https://renderio.dev/api/v1/run-ffmpeg-command \
  -H "Content-Type: application/json" \
  -H "X-API-KEY: your_api_key" \
  -d '{
    "ffmpeg_command": "-i {{in_video}} -vf \"crop=iw-6:ih-6:3:3,noise=alls=6:allf=t,eq=brightness=0.015:saturation=1.02,hue=h=1\" -af \"asetrate=44100*1.003,aresample=44100\" -c:v libx264 -profile:v high -level:v 4.0 -crf 23 -preset medium -pix_fmt yuv420p -c:a aac -b:a 128k -ar 44100 -map_metadata -1 -fflags +bitexact -movflags +faststart {{out_video}}",
    "input_files": { "in_video": "https://example.com/ai-video.mp4" },
    "output_files": { "out_video": "naturalized.mp4" }
  }'
Enter fullscreen mode Exit fullscreen mode

Tool-specific considerations

Runway Gen-3/Gen-4

Runway writes multiple custom metadata fields: encoder, handler_name, and sometimes a comment field with generation parameters. The -map_metadata -1 -fflags +bitexact flags strip all of these.

Runway's color profile tends toward high saturation with punchy contrast. The naturalizer command's brightness and saturation shifts already handle this, but if your video still looks "too clean," add a slight gamma adjustment: gamma=0.98 in the eq filter.

Runway Gen-4 outputs at exactly 24fps with zero frame timing variation. Real phone cameras shoot at 29.97 or 30fps with slight jitter. Re-encode at 30fps with the timing variation from Step 3.

Kling AI

Kling has a known issue with temporal inconsistencies at scene transitions — frames sometimes stutter or repeat. The noise filter masks these, but also check for it visually before uploading. A single repeated frame is a dead giveaway to human reviewers.

Kling may embed watermarks depending on your subscription tier. Check the bottom-right corner of the frame. If present, crop by 20-30 pixels from the bottom edge:

ffmpeg -i kling_video.mp4 -vf "crop=iw:ih-30:0:0" -c:v libx264 -crf 22 output.mp4
Enter fullscreen mode Exit fullscreen mode

Kling's audio tracks are often silent or contain synthesized ambient noise at suspiciously consistent levels. If your video has audio, verify it sounds natural or replace it entirely.

Sora

Sora produces some of the smoothest AI video on the market, which is actually a problem. Real video has micro-jitter, slight focus shifts, and compression artifacts from the camera sensor. Sora has none of that.

Beyond the noise and timing variation from the naturalizer command, consider adding a slight speed fluctuation. Slow the video by 2% and it introduces natural-feeling drag:

ffmpeg -i sora_video.mp4 -vf "setpts=PTS*1.02,noise=alls=7:allf=t" -c:v libx264 -crf 22 output.mp4
Enter fullscreen mode Exit fullscreen mode

Sora also outputs with specific C2PA Content Credentials that explicitly declare AI generation. The metadata strip handles this, but double-check with ffprobe after processing.

Pika Labs

Pika's free tier adds a visible watermark in the lower-right corner. Crop it or cover it with your own overlay before running the naturalizer:

ffmpeg -i pika_video.mp4 -vf "crop=iw-40:ih-40:0:0" -c:v libx264 -crf 22 output.mp4
Enter fullscreen mode Exit fullscreen mode

Pika's output resolution varies by model version (sometimes 576p, sometimes 720p, sometimes 1080p). If you're uploading to TikTok, resize to 1080x1920 after naturalizing. A non-standard resolution is a subtle signal.

For more on cleaning up AI artifacts specifically, see remove AI artifacts from video and make AI video look natural.

Batch processing AI videos

For content operations generating many AI videos:



API_KEY = "ffsk_your_key"
HEADERS = {"Content-Type": "application/json", "X-API-KEY": API_KEY}

videos = [
    "https://example.com/ai-video-1.mp4",
    "https://example.com/ai-video-2.mp4",
    "https://example.com/ai-video-3.mp4",
]

for i, url in enumerate(videos):
    noise = 5 + (i % 4)
    brightness = 0.01 + (i * 0.005)

    response = requests.post(
        "https://renderio.dev/api/v1/run-ffmpeg-command",
        headers=HEADERS,
        json={
            "ffmpeg_command": f'-i {{in_video}} -vf "crop=iw-6:ih-6:3:3,noise=alls={noise}:allf=t,eq=brightness={brightness:.3f}" -c:v libx264 -crf 23 -map_metadata -1 -fflags +bitexact {{out_video}}',
            "input_files": {"in_video": url},
            "output_files": {"out_video": f"natural_{i}.mp4"}
        }
    )
    print(f"Video {i}: {response.json()['command_id']}")
Enter fullscreen mode Exit fullscreen mode

Verification

After processing, check four things:

Run ffprobe -v quiet -print_format json -show_format output.mp4 and confirm no tool-specific metadata fields remain. Look for encoder, comment, creation_time, and any custom tags. If any are present, your metadata strip didn't work.

View the video at 200% zoom. You should see subtle grain from the noise filter. If the image is perfectly clean, the noise wasn't applied.

Check file size. A 30-second 1080p video from a phone is typically 30-80MB. If your output is 5MB or 200MB, something's off with the encoding settings.

Play the audio back. A 0.3-0.5% pitch shift is inaudible. Above 1%, you'll hear it. If the audio sounds slightly chipmunked, dial back the pitch multiplier.

If you're posting the same AI video to multiple accounts, you'll also need to avoid TikTok duplicate detection at scale by generating unique variations. For deeper metadata removal, the remove AI metadata from video guide covers edge cases that the basic strip misses.

Get started

The Starter plan at $9/mo includes 500 commands, enough to process 10-15 AI-generated clips per day. Explore the FFmpeg video API or get your API key.

FAQ

Does TikTok actually detect AI-generated video?

Yes, and it's getting better at it. TikTok uses a combination of metadata analysis, perceptual fingerprinting, and (increasingly) visual pattern detection. C2PA Content Credentials are the most obvious signal. Tools like Runway and Sora now embed these by default. Metadata stripping handles C2PA. The visual patterns are harder to detect algorithmically, but TikTok is investing in it.

Will these techniques work on Instagram and YouTube too?

The same principles apply. Instagram uses similar fingerprinting for Reels. YouTube has its own content detection system (Content ID) but it's focused on copyright, not AI detection, at least for now. The metadata strip and noise addition work across all platforms.

Is it legal to remove AI metadata from videos?

Removing metadata is legal in most jurisdictions. However, some regions are implementing AI disclosure requirements (the EU AI Act, for example). Removing C2PA markers to avoid disclosure could have legal implications depending on how you use the video. This guide covers the technical steps; consult local regulations for compliance.

How much noise should I add without making the video look bad?

Noise strength 5-8 is the range. Below 5, the noise is too subtle to fool detection. Above 10, it's visible on mobile screens. For high-quality AI video (Sora, Runway Gen-4), start at 6. For lower-quality sources (Pika free tier, older Kling models), start at 4. They already have enough imperfections.

Do I need to process audio separately?

Not usually. The naturalizer command shifts audio pitch as part of the combined pipeline. If your AI video has no audio (many AI tools generate silent video), add a natural ambient track or keep it silent. TikTok doesn't flag silent videos specifically. If you're adding music, that replaces the audio fingerprint entirely.

Top comments (0)