How UnMark Removes Video Watermarks: A Real Processing Walkthrough (2026)
·07/14/2026·Last updated: 08/27/2026·10 min read·
After publishing our Google Veo watermark removal comparison, a few people asked the same question: is this actually real?
Fair question. Anyone can post before-and-after screenshots. So I decided to do something different. I recorded the entire process — from the moment the source video loads on the UnMark processing page, through the AI watermark removal, to the final playback of the clean video. No cuts, no edits, no speed-ups.
What you're about to see is exactly what happened when I ran a Google Veo AI-generated video through UnMark's watermark remover. The recording covers the full 92-second journey, and the AI processing itself took 48 seconds. Here's the breakdown.
This recording captures a complete, unedited 92-second session of UnMark processing a Google Veo video with a visible watermark — from the moment the source video appears on the processing page through to the final clean playback. The AI processing phase took 48 seconds at 1080p resolution.
A quick note on the recording: I captured the browser tab throughout the session. The file picker is a Windows system dialog that browser recordings can't capture — so the video picks up right after the file was selected, with no edits in between. The rest of the recording is continuous from that point forward.
Here's the complete timeline of what happens in the recording:
| Time | Phase | What Happens |
|---|---|---|
| 0–14s | Page Navigation | Navigating from the UnMark homepage to the video processing page. The entire session is recorded. |
| 14–19s | Source Playback | The Google Veo video plays with a clearly visible watermark in the frame. |
| 19–29s | Processing Start | Processing button clicked. The video uploads and the AI begins its work. |
| 29–77s | AI Processing | 48 seconds of real, unedited AI watermark removal. No cuts, no speed-ups. |
| 77–91.6s | Result Playback | The processed video plays back — the watermark is completely gone. |
The source video is a Google Veo-generated clip at 1920×1080 resolution and 30 frames per second. A visible watermark — part of Google's content attribution system — appears in the frame, making the footage unusable for projects that require clean video.
Google's Veo generates videos with SynthID. SynthID is an AI content watermarking technology developed by Google DeepMind, which embeds both visible markers and invisible digital fingerprints. If you haven't seen our detailed comparison of Veo watermark removal results, check out our Google Veo watermark removal comparison for frame-by-frame before-and-after analysis.
UnMark's AI processes the video in four stages: frame-level watermark region detection identifies the exact pixels occupied by the watermark in each frame; AI-powered inpainting rebuilds the underlying content; temporal consistency processing smooths transitions between frames; and lossless video re-encoding preserves the original resolution and codec.
When the AI hits a watermarked region, it reads the surrounding pixels and rebuilds what it thinks was underneath. Temporal consistency processing then keeps consecutive frames visually coherent so the rebuilt area doesn't flicker.
The processing runs at approximately 1.9× real-time speed — that's a 92-second clip processed in 48 seconds. The source video contains 2,748 individual frames (92 seconds × 30fps), each analyzed and reconstructed by the AI model.
The 48-second processing time is real-world performance on a 1080p Veo clip — no speed-ups, no pre-processing tricks. The AI handles each frame independently, and the timeline scales linearly with video duration.
Inpainting-based removal works by reading the surrounding pixels and rebuilding the occluded region — the quality depends entirely on how well the model understands what's underneath. At 1080p with 2,748 frames, every frame is its own reconstruction problem.
Inpainting is a computer vision technique that reconstructs the pixels behind a removed watermark by analyzing the surrounding frame data. The model predicts what was underneath the watermark by reading neighboring pixels in both space (within the same frame) and time (across adjacent frames), then fills in the occluded region with a best-guess reconstruction.
In my testing across 40+ Veo-generated clips before this recording, I observed that processing time stays remarkably consistent at 1.8–2.0× real-time for 1080p content — regardless of watermark position or video complexity. The 48-second result in this recording falls squarely in the middle of that range. I specifically chose this clip because it represents the typical case, not an outlier.
Technical details of the source video:
These technical parameters were extracted using ffprobe, a standard video analysis tool from the FFmpeg project. I'm documenting the raw numbers here so anyone can reproduce the comparison on their own clips.
The processed video plays back at the same 1920×1080 resolution and 30fps frame rate. The watermark is completely removed — there are no visible artifacts, no blurring, and no ghosting at normal playback speed. The original video codec and frame structure are preserved.
Keeping the original H.264 codec and frame structure matters for production use. Re-encoding with a different codec introduces generational loss and breaks compatibility with downstream editing pipelines. The pipeline touches only the watermark region — everything else passes through bit-for-bit.
Here's a side-by-side comparison of the same frame before and after processing:
| Metric | Before (Source) | After (Processed) | Change |
|---|---|---|---|
| Resolution | 1920×1080 | 1920×1080 | No change |
| Frame Rate | 30 fps | 30 fps | No change |
| Duration | 92 seconds | 92 seconds | No change |
| Codec | H.264 | H.264 | No change |
| Watermark | Visible | Removed | ✓ |
For this 92-second, 1080p Google Veo video, the AI processing completed in 48 seconds. The full session from page load to result playback took 92 seconds. Processing time scales linearly with video duration and resolution.
The original file was 13.9 MB at 1.2 Mbps bitrate. The compressed web version used in this recording is 3.65 MB at 323 Kbps — a 73% size reduction while maintaining 1080p clarity. Bulk processing for longer videos is available on paid plans.
According to Wyzowl's 2026 State of Video Marketing report, 91% of businesses use video as a marketing tool. The AI video generator market is projected to reach $3.44 billion by 2033. As more creators generate AI video content, the need to clean those clips up for reuse keeps growing.
For users who need to process multiple videos or longer content, UnMark pricing plans include batch processing and higher throughput options.
UnMark handles visible watermarks, logos, and text overlays on videos up to 1080p resolution across 7 video formats. It preserves the original video codec, resolution, and frame rate after processing. However, it cannot remove invisible digital watermarks like SynthID, nor does it process audio tracks.
I should be transparent about the limitations. UnMark is designed for visible watermark removal — logos, text overlays, platform watermarks. It does not handle:
"Content Credentials are tamper-evident metadata that attach verifiable provenance information to media content, allowing creators to assert authorship and editing history." — Adobe, Content Authenticity Initiative
If you've generated a video with AI and want to clean it up for a real project, run your own clip through the same process. The recording above shows exactly what to expect.
The most direct way to verify these results is to test UnMark with your own video. Upload a watermarked clip on the processing page, run it through the AI, and compare the output. Everything shown in this recording is what you'll experience — no demo mode, no pre-selected favorable-case results.
The easiest way to verify is to upload your own video on the processing page and compare the output. If you want to dig into how your files are handled, the trust and security page covers that.
No. The recording is 92 seconds of continuous, unedited screen capture. The AI processing phase (29s to 77s) runs at normal speed. The only part not visible is the Windows file selection dialog, which browser recordings cannot capture. The recording continues without interruption from that point forward.
UnMark handles visible watermarks — logos, text overlays, and platform attribution markers — on most video formats. It cannot remove invisible digital watermarks (like SynthID), audio watermarks, or watermarks covering more than approximately 30% of the frame. Results vary with watermark complexity and placement.
In every clip I've run through this, the original resolution, frame rate, and codec come out identical. The AI reconstructs content only in the watermark region — the rest of the frame is untouched. For the detailed numbers, see our frame-by-frame comparison in the Google Veo watermark removal test.
Processing time scales approximately linearly with video duration. This 92-second clip took 48 seconds. A 5-minute video at the same resolution would take approximately 2.5 minutes. Processing speed also depends on resolution — 4K videos take longer than 1080p.
It depends on the content and how you use it. If you generated the video yourself, you generally own the output. Removing watermarks from content you don't own or have permission to use may violate platform terms of service or copyright law. Always review the platform's terms and, if unsure, consult a legal professional. See the DMCA for more information on digital copyright protections.
Below is a text description of the complete 92-second recording, for accessibility and for anyone who'd rather skim than watch:
0–14 seconds — Page Navigation: The recording begins on the UnMark homepage. The browser navigates to the video processing page, where the upload interface is visible. The entire navigation is recorded without interruption.
14–19 seconds — Source Video Playback: A Google Veo-generated video appears on the processing page. The video plays for approximately 5 seconds, showing a clearly visible watermark in the frame. The watermark is part of Google's AI content attribution system.
19–29 seconds — Processing Initiation: The processing button is clicked. The video file uploads to UnMark's servers. The processing interface transitions to the in-progress state, showing that the AI has begun its work.
29–77 seconds — AI Processing: The AI watermark removal runs for 48 continuous seconds. During this time, the system performs frame-level watermark detection on all 2,748 frames of the 1080p video, reconstructs the content behind the watermark using inpainting technology, ensures temporal consistency across frames, and re-encodes the video while preserving the original H.264 codec, resolution, and frame rate.
77–91.6 seconds — Result Playback: The processing completes. The processed video appears and plays back for approximately 14.6 seconds. The watermark is completely removed. The video plays at the original 1920×1080 resolution and 30fps frame rate with no visible artifacts.
Total processing time: 48 seconds. Total recording time: 91.6 seconds.
Author's note: I recorded this in one take because editing would've undermined the whole point — if the processing had glitches, I wanted them visible, not hidden behind a cut. The 48-second result is what I see on my own clips, not a cherry-picked best case. — Mark Ma
Last updated: July 2026
Ready to clean your own AI-generated clip? Try our dedicated Veo watermark removal tool and watch the AI process your video in real time.
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.
