Mark Ma is the founder and lead engineer at UnMark, where he leads the development of AI-powered video watermark removal technology. With over a decade of experience in computer vision, deep learning, and video processing, he has built production systems that process millions of frames per month across TikTok, Instagram Reels, and Google Veo content. His technical work focuses on motion-compensated watermark detection, spatial-temporal inpainting, and quality-preserving reconstruction for short-form vertical video. Before UnMark, Mark shipped computer vision infrastructure at scale for content moderation and ad compliance platforms. He writes this blog to document the real engineering, testing, and regulatory reasoning behind UnMark's processing pipeline — every test result, frame analysis, and benchmark published here is reproduced from actual production data, not synthetic examples. His goal is to give creators, brands, and regulators a verifiable technical reference for AI video watermark removal in 2026.
Real test of watermark removal on a Google Veo AI-generated video. Frame-by-frame comparison with actual data: resolution, bitrate, and quality before/after.
Facebook names a bookmark as if it were a download. Native lane, the Save trap, five URL forms, and how to fetch clean MP4s from any public video.
X gates the download button by subscription, not watermark. Premium lane vs link lane, why X file sizes are so often wrong, and how to fetch clean MP4s from any account.
Instagram's save button is a bookmark, not a file - and the export stamps a watermark. Fetch the clean CDN source instead, and learn which of three watermark types you are actually facing.
Skip the branded save: TikTok's play_addr stream is the clean original the app plays. Copy the link, parse free, take the MP4 - no watermark ever touches the file.
Downloading without removing: fetch the clean source file platforms use for playback. 15 verified platforms, seconds per clip - plus the decision frame for when AI removal is still needed.
How watermark detection models work: CNN detection, SAM 2 mask generation, temporal tracking for moving marks, and the invisible-watermark research frontier.
DMCA Section 1202 makes watermark stripping its own violation: CMI rules, per-work damages to $25,000, the 2026 scienter ruling, and when removal is legal.
Sora's logo cycles 3 regions every ~3 seconds. Tested late-2026 workflow to clean it with motion-aware AI inpainting while C2PA provenance stays intact.
Facebook burns two overlays into every downloaded Reel. Tested 2026 workflow to remove the FB logo and handle with AI inpainting — plus what Meta Content Protection means for compliance.
YouTube burns an @handle watermark into every downloaded Short. Tested 2026 workflow to remove it with AI inpainting and keep your 9:16 frame intact.
Batch video watermark removal turns a weekend chore into a 30-minute background task. Tested workflow for cleaning up to 20 videos at once in 2026.
