Last week I found myself staring at a Google Veo-generated video, wondering the same thing a lot of creators are probably wondering right now: what happens if you try to remove the watermark from an AI-generated video?

So I did what any engineer would do. I ran it through our watermark remover and documented every single thing that happened. Frame by frame. Bitrate by bitrate. No marketing fluff — just the actual results.

Here's what I found.

Key Findings

Resolution & frame rate: 100% preserved (720×1280, 24fps, H.264)
Visual quality: Watermark removed cleanly at all 5 test timestamps (1s, 2.5s, 4s, 5.5s, 7s)
Video bitrate: Minor increase from 1.62 Mbps to 1.75 Mbps (+8.1%) due to AI content reconstruction
Audio: Completely untouched (AAC, 72 Kbps, identical in both files)
File size: Grew from 1.62 MB to 1.74 MB (+7.7%) — negligible for an 8-second clip
Bottom line: Watermark removal on Google Veo AI video works without quality degradation

Google Veo AI-generated video before watermark removal — Before — with Veo watermark

Google Veo video after UnMark watermark removal — After — watermark removed

What Is Google Veo — and Why Should You Care?

Google Veo is Google's generative video model. You give it a text prompt, and it spits out a video. Think DALL-E or Midjourney, but for moving images. Veo was first announced at Google I/O in May 2024, and the Veo 2 model followed in December 2024, pushing the quality bar significantly higher.

Here's what makes Veo different from the AI image generators most people are familiar with: it doesn't just generate a single frame — it has to maintain consistency across dozens or hundreds of frames. That's a fundamentally harder problem, and it's why video generation has lagged behind image generation by a couple of years.

According to Wyzowl's 2026 State of Video Marketing report, 91% of businesses use video as a marketing tool — back to joint all-time highs after a slight dip in 2025. As AI video generators like Veo, Sora, and Runway Gen-3 become more accessible, that number is only going to go up. Grand View Research values the AI video generator market at $788.5 million in 2025 and projects it will reach $3.44 billion by 2033, growing at a 20.3% compound annual rate.

But there's a catch. Most AI-generated videos come with some kind of visual marker — either a platform logo, a "Generated by AI" label, or a digital watermark like Google's SynthID. If you want to use that AI-generated footage in a real project, those markers are a problem.

The Watermark Problem in AI-Generated Videos

When AI generates a video, the platform usually adds some form of attribution. Sometimes it's obvious — a logo in the corner. Sometimes it's subtle — an invisible digital watermark embedded in the pixels. Google's SynthID, for example, embeds a digital watermark directly into the video frames that's designed to survive compression, screenshots, and even partial cropping.

For anyone working with video processing, this creates a real dilemma for creators. You've got this impressive AI-generated footage, but it's branded. You can't use it in a client project with a watermark. You can't mix it with your own footage without the watermark being a distraction. And cropping it out — the old-school approach — means losing part of your frame.

"SynthID adds an invisible digital watermark to an AI-generated image (or video segment). The watermark doesn't change the image or video quality. It's added the moment content is created, and designed to stand up to modifications like cropping, adding filters, changing frame rates, or lossy compression." — Google DeepMind, SynthID official documentation

So the question became: what actually happens when you try to remove it?

Test Methodology: Setup and Metrics

The goal here was simple: take a Google Veo-generated video, remove the watermark, and then compare the original and the processed version side by side to see if anything changed.

Here's the setup:

Source video: A Google Veo-generated clip, 720×1280 resolution (vertical format — think TikTok or Reels), 24 frames per second, running exactly 8 seconds. The file was encoded in H.264 at 1.62 Mbps, with AAC audio at 72 Kbps. Total file size: 1.62 MB.
Tool used: UnMark's AI watermark removal, Quality mode
What was measured: Resolution, frame rate, bitrate, file size, and — most importantly — visual quality through frame-by-frame comparison

The exact technical specs were pulled from both files using ffprobe. No estimates, no rounding — these are the actual numbers from the video files.

Original Veo video frame at 1 second showing watermark

Frame-by-Frame Visual Comparison Results

Frames were extracted at five different timestamps — 1 second, 2.5 seconds, 4 seconds, 5.5 seconds, and 7 seconds — from both the original and the processed video, then lined up side by side.

At the 1-second mark, the original video shows the watermark clearly. The processed version? No trace of it. The background — a complex gradient with moving elements — is clean. No blurring, no smudging, no obvious patch jobs. Without seeing the original, it would be impossible to tell there was ever a watermark there.

Veo video frame at 1 second before processing — Before (1s)

Veo video frame at 1 second after watermark removal — After (1s)

At 2.5 seconds, the scene has more motion. This is where older watermark removal methods usually fall apart — you get ghosting, flickering, or visible seams. The processed video here is smooth. The AI's temporal consistency handling (basically, making sure the reconstructed content doesn't jump around between frames) is holding up well.

Veo video frame at 2.5 seconds before processing — Before (2.5s)

Veo video frame at 2.5 seconds after watermark removal — After (2.5s)

At 4 seconds, we hit the middle of the clip. The watermark sits right over a detailed area of the frame — not just a flat background, but actual content. This is the hardest kind of removal to pull off. The result: the content underneath is reconstructed. Is it pixel-perfect? Not quite. But at normal playback speed, you genuinely can't tell. You'd need to pause and zoom in to find any difference.

Veo video frame at 4 seconds before processing — Before (4s)

Veo video frame at 4 seconds after watermark removal — After (4s)

At 5.5 seconds and 7 seconds, the same story. Clean removal, consistent quality, no artifacts that jump out at normal viewing speed.

Veo video frame at 5.5 seconds before processing — Before (5.5s)

Veo video frame at 5.5 seconds after watermark removal — After (5.5s)

Veo video frame at 7 seconds before processing — Before (7s)

Veo video frame at 7 seconds after watermark removal — After (7s)

Technical Metrics: Resolution, Bitrate, and Quality

This is where the data gets interesting. Both files were analyzed with ffprobe. Here's what came back:

Metric	Original (Veo)	Processed	Change
Resolution	720 × 1280	720 × 1280	No change
Frame Rate	24 fps	24 fps	No change
Duration	8.0 seconds	8.0 seconds	No change
Video Codec	H.264	H.264	No change
Audio Codec	AAC	AAC	No change
Video Bitrate	1.62 Mbps	1.75 Mbps	+8.1%
Audio Bitrate	72 Kbps	72 Kbps	No change
File Size	1.62 MB	1.74 MB	+7.7%

Three things stand out in the data.

First, zero quality loss. The resolution, frame rate, and codec are identical. The processing didn't downscale or re-compress the video in a way that degrades quality. If you've ever used video editing tools that re-encode your footage and leave it looking worse than the original, you know how rare this is.

Second, the bitrate actually went up slightly. From 1.62 Mbps to 1.75 Mbps — about an 8% increase. This isn't a bug. When the AI reconstructs the content underneath the watermark, it's generating new pixel data. That new data has to be encoded, and encoding it at the same quality level requires slightly more bits than the original compressed version. The file grew from 1.62 MB to 1.74 MB. For an 8-second clip, that's negligible. For a 10-minute video, you'd be looking at roughly 7-8% more storage.

Third, audio is untouched. The AAC audio track at 72 Kbps is identical in both files. No re-encoding, no quality loss. The AI only processes the video frames — the audio stream passes through unchanged.

"The encoding (compression) process is typically lossy — it degrades stream quality to make the output smaller." — FFmpeg official documentation. This is why preserving the original resolution, frame rate, and codec matters: every re-encode generation introduces quality loss that compounds. A tool that downscales your 1080p video to 720p destroys over 56% of the pixel data, and most users never notice until the footage is already in a client deliverable.

Unexpected Findings and Analysis

The expectation going into this test was to find some kind of tradeoff. Better watermark removal usually means worse quality — that's been the rule for years. Crop it out and you lose the edges. Blur it and you get a smudge. Clone-stamp it in a video editor and you'll spend 20 minutes per frame.

The surprise wasn't that the watermark was gone — that was expected. The real finding was that the rest of the video was completely unchanged. Same resolution, same frame rate, same audio. The only thing the AI touched was the watermark region itself.

The bitrate increase was also unexpected. Intuitively, removing something should make the file smaller, not larger. But on closer inspection, it makes sense — the AI is generating new visual information where the watermark used to be, and that information has to be encoded. The original watermark pixels were simple, predictable shapes (logos and text). The reconstructed content is complex, natural-looking imagery. Complex images take more bits to encode than simple logos.

One important caveat: this was a relatively short clip — 8 seconds. For longer videos, processing time scales roughly linearly with duration. A 1-minute video at the same resolution would take proportionally longer, but the per-frame quality should be consistent.

What This Means for Content Creators

If you're a content creator experimenting with AI video generation — whether it's Google Veo, OpenAI's Sora, Runway, or any of the other tools flooding the market — here's the honest take:

AI-generated footage is genuinely useful. It's not going to replace your camera anytime soon, but for B-roll, background plates, concept visualization, and social media content, it's already good enough. The problem is the watermark. If you're embedding AI-generated clips in a larger project — a YouTube video, a client deliverable, a presentation — that watermark is a dealbreaker.

Based on this test, watermark removal on AI-generated video works. The quality holds up. The technical specs hold up. If you're wondering whether it's worth trying on your own Veo or Sora clips, the data says yes.

That said, a few things to keep in mind:

Ownership matters. You should only remove watermarks from content you generated yourself or have explicit permission to process. Google's terms of service for Veo are worth reading carefully — different platforms have different rules.
Invisible watermarks are a separate issue. Google's SynthID embeds a digital fingerprint that survives visual edits. Removing a visible watermark doesn't remove SynthID. If you need to prove your content is AI-generated for transparency reasons, the invisible watermark remains intact.
Test with your own footage. Every AI video generator has slightly different output characteristics. What worked for this Veo clip might behave differently with Sora or Runway footage. Run your own tests.

FAQ

Does removing the watermark affect video quality?

Based on this test: no. Resolution, frame rate, and codec were identical between the original and processed files. The video bitrate increased slightly (+8.1%), which is expected when the AI reconstructs content in the watermark region.

Can you remove invisible watermarks like SynthID?

No. Google's SynthID is a digital watermark embedded at the pixel level — it's designed to survive visual edits, compression, and format conversion. Removing a visible watermark does not affect SynthID or similar invisible watermarking technologies.

How long does the processing take?

For this 8-second, 720p clip, processing took a few minutes in Quality mode. Processing time scales with video duration and resolution. A 1-minute 1080p video would take proportionally longer.

Is it legal to remove watermarks from AI-generated videos?

It depends on the platform's terms of service and how you're using the content. If you generated the video yourself using a tool like Veo, you generally own the output (check the specific terms). Removing the watermark from your own content is different from removing it from someone else's. Always review the platform's terms and, if you're unsure, consult a legal professional.

What AI video generators was this tested with?

This specific test used a Google Veo-generated video. The principles should apply to other AI video generators (Sora, Runway, Pika, etc.), but those haven't been tested yet. The watermark style, placement, and opacity vary between platforms, which can affect removal quality.

Does the processed video look exactly like the original?

At the watermark region, the AI reconstructs what was underneath — so the processed version is slightly different from what the original would look like without a watermark, since the reconstruction is a prediction, not a restoration. At normal playback speed, the difference is imperceptible. Frame-by-frame, pixel-peeping at 400% zoom, you might spot minor differences in the reconstructed area.

Bottom Line

This test ran a Google Veo AI-generated video through UnMark's watermark remover. The watermark came off clean. The resolution, frame rate, and audio were identical. The bitrate went up slightly — from 1.62 to 1.75 Mbps — because the AI-reconstructed content is more complex than the original watermark pixels. The file grew from 1.62 MB to 1.74 MB, which is negligible for an 8-second clip.

If you're using AI video generators and want to use the footage in actual projects, try UnMark's watermark remover — this approach works. The data backs it up. Just make sure you own the content or have permission to process it, and — as always — test with your own footage before committing to a workflow.

I'll be testing this with Sora and Runway clips next. If you've tried something similar, I'd love to hear about it.