Hey everyone,
When you generate images using Google’s AI tools (like Gemini), they automatically add an invisible watermark called SynthID that tags the image as AI-generated.
I’m looking for a way to remove or bypass this watermark so the image doesn’t get flagged as AI-made.
Anyone know how to do this?
If you look through my profile there are links to Christmas give away software…….. I just downloaded a program yesterday that removed watermarks
I don’t see any software that can remove SynthID.
SynthID is Google’s invisible pixel-level watermark (plus that annoying visible diamond/logo in Gemini/Nano Banana images), and it’s built to survive most basic edits like cropping, resizing, compression, or filters. Google says it’s pretty robust on purpose for transparency.
That said, in 2026 the community has figured out some solid workarounds that actually work for a lot of people:
For the visible logo/diamond (quick & easy):
Just crop it out (even if it messes the ratio a tiny bit).
Or use free tools like Google Photos Magic Eraser, or the open-source GitHub repo GargantuaX/gemini-watermark-remover (pure JS, client-side, uses reverse alpha blending – lossless and precise).
For the actual invisible SynthID:
Best free option right now: Upload your image to https://www.gptcleanup.com/gemini-image-watermark-remover – it runs 100% in your browser (no upload to servers), uses imperceptible pixel perturbations, and strips the SynthID marker.
Another one that claims good results: https://www.chromastudio.ai/synthid-remover-image – AI scans and erases it while keeping quality.
More advanced/open-source route: Check the ComfyUI “SynthID-Bypass V2” project or the recent Medium guide on “spectral bypass” methods (they reduce the watermark energy so detectors fail the confidence threshold).
These aren’t 100% foolproof every single time (success rates float around 70-90% depending on the image and how much you edit afterward), and heavy compression or re-uploading through another generator can help further scramble it. Google officially hates this and their terms say don’t bypass transparency markers, so use at your own risk.
If you’re just testing or for personal stuff it’s whatever, but don’t use it for anything deceptive – detectors are getting better and ethics matter.
Anyone got newer/better methods that are working in 2026? Drop 'em below!
Right, so what you’re chasing:
Reads like you hit a wall of identical “SynthID Remover” sites and bounced over for the real answer. Fair enough — that wall is where everyone gets stuck.
Here’s the kick:
on Instagram, Pinterest, and Facebook — SynthID isn’t even what’s flagging you.
You know how every photo carries hidden tags like camera model, GPS, date taken?
Instagram reads one specific kind of that called C2PA Content Credentials to slap the “Made with AI” label on your post.
SynthID is a different beast — invisible noise woven into the pixels themselves.
Platforms read one. Google’s own detector reads the other. Different doors, different keys.
The full stack of what’s actually checking your images:
| Layer | What flags it | How to handle | |
|---|---|---|---|
 |
C2PA metadata | Instagram, Pinterest, Facebook “Made with AI” labels | exiftool -all= → 10 seconds, free, done |
 |
SynthID pixel watermark | Google’s own SynthID Detector | Hard mode — ComfyUI, 16–79% per method |
 |
Platform visual classifier | Stock sites (Adobe, Shutterstock, Getty) | Source quality matters more than tricks |
 |
Google neural-hash DB (suspected) | Google’s internal systems only | Unbeatable — only counter is don’t use Google |
For 90% of “I don’t want my image labelled AI on Instagram” — you only need layer 1.
Run it through exiftool on the command line, or aimetadatacleaner.com in your browser (stays on your device, nothing uploads).
alias clean='exiftool -all= -overwrite_original'
clean your-image.png
I keep that alias on my Mac and run it on anything before posting. Caught a “Made with AI” label twice last month from Gemini images I’d only lightly touched up in Photoshop. Heads up: older exiftool versions (12.x) sometimes miss the embedded C2PA manifest — mine on 13.30 handles JUMBF blocks fine. Upgrade, or use stripshot.app as a fallback.
The lazier 100% solution: generate with FLUX, Stable Diffusion, or Midjourney instead. None of them embed SynthID. The simplest bypass is not stepping into the trap.
This only matters if the thing checking your image runs Google’s SynthID Detector specifically — which today is basically just Google’s own apps. For Instagram and friends, you’re already done above.
If you need it anyway, three working setups exist as of right now (top two updated within the last 2 weeks):
| Tool | Strength | Catch |
|---|---|---|
00quebec/Synthid-Bypass v2 → drag Synthid-Bypass-v2.0.json onto your ComfyUI canvas |
Best face preservation. Qwen + Canny ControlNet + face-detailer pass. | Needs 16GB+ VRAM. |
aloshdenny/reverse-SynthID Round 06 → python scripts/dissolve_batch.py --strengths final nuke |
Tested directly against gemini-3.1-flash-image-preview and nano-banana-pro-preview. Spectral subtraction + 7-stage stack. |
Per-model codebook required. |
mertizci/noai-watermark → brew install then one CLI command |
Easiest install. CtrlRegen-based. | Less polished, works in most cases. |
Honesty check before you commit hours. Aloshdenny — the guy who built reverse-SynthID, after weeks of work and 123k image pairs — writes:
“It’s impossible to remove SynthID 100%, but it is POSSIBLE to remove portions of it, to a point where detection fails.”
His V2 hit only ~16% evasion. Round 06 stacks 7 transformations to climb back near 100%. Single-method attacks (just JPEG, just img2img, just noise) embarrassingly fail. The SaaS removers on Google page 1? Nobody shows receipts.
If your first run on a portrait makes the face look slightly off — that’s normal, not a fail. Drop denoise from 0.2 down to 0.15, or bump the face-detailer strength. Run it again.
The whole field is week-by-week. What defeats today’s Gemini detector might not survive the next quiet patch.
Awesome-GenAI-Watermarking — catalogs every published watermarking scheme and every documented attack against it. When the current methods get patched, the next batch lands here first.
Also worth knowing:
You came in chasing one watermark.
Walking out with the full map — and the realisation that three of the four layers are either a one-line fix or a “use a different model” sidestep.
The hunt is over.
!