remove-ai-watermarks: a Python CLI that splits AI mark removal into GPU and non-GPU jobs
AI watermark remover. CLI and Python library to strip visible and invisible AI watermarks (Gemini / Nano Banana sparkle, SynthID) and provenance metadata (C2PA, EXIF, IPTC) from images.
At a glance
- What is it?
- wiltodelta/remove-ai-watermarks is an Apache-2.0 Python CLI and library for stripping visible AI labels, invisible pixel watermarks and provenance metadata from your own images and video. Metadata and visible marks run on CPU; invisible removal needs CUDA.
- Who is it for?
- Use it if you own the images or video and need to clear Gemini-style visible marks or C2PA, EXIF and IPTC metadata without a GPU. Skip it for text, for other people's paid content, or if you expect a local SynthID detector.
- Can I use it commercially?
- Yes. Apache-2.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository received new commits within the last day.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- wiltodelta/remove-ai-watermarks6,401 ★ · +2,076
| Repository | 2026-07-28 | 2026-11-15 |
|---|---|---|
| wiltodelta/remove-ai-watermarks | 4.3k | 6.4k |
2,076 Stars in the last 110 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
Fourteen subcommands, sorted by whether they need a GPU
The command surface is organized by goal rather than by file type. identify finds provenance signals and watermarks. classify judges a photograph from its pixels. visible removes known visible AI marks, erase removes a region you select, and metadata strips AI metadata. For video there is video identify, video all, video metadata, video visible, video batch and video invisible. Finally there are invisible, all and batch for the heavier image paths.
The GPU column is the part that decides what you can do on a given machine. identify, classify, visible, erase, metadata, video identify, video all, video metadata and video visible are all marked as needing no GPU. video batch and batch depend on the mode you choose. video invisible and all list a GPU as recommended. Only invisible lists CUDA as required.
The practical consequence is that a laptop without an NVIDIA card can still do the common cleanup jobs: clear a visible label, drop C2PA, EXIF, XMP and IPTC fields, and process video for the registered visible marks. The registered video marks are Sora, Veo, Seedance, Doubao, Dola, Hailuo AI and Kling AI. What the same laptop cannot do is the image regeneration route that targets invisible watermarks. If a project needs that, you either find a CUDA machine or use the hosted service at raiw.cc.
The default install only handles metadata, and every other job is an extra
The plain package is the metadata-focused CLI. Visible removal, classification, video and diffusion each sit behind a separate extra, so the install command has to match the job you intend to run.
uv tool install remove-ai-watermarksThe extras table maps needs to installs. Metadata inspection and stripping is the bare remove-ai-watermarks package. Photograph AI-versus-camera classification is remove-ai-watermarks[classify]. The 10-provider source-export and model-version classifier is remove-ai-watermarks[source-classify]. Visible detection and removal is remove-ai-watermarks[visible], visible video processing is remove-ai-watermarks[video], and video SynthID removal is remove-ai-watermarks[video,diffusion]. Torch-free DWT-DCT detection uses remove-ai-watermarks[detect], direct local Paint InvisMark disruption uses remove-ai-watermarks[pixels], and invisible image removal that needs CUDA uses remove-ai-watermarks[qwen-zimage]. The catch-all remove-ai-watermarks[all] covers every production feature available on the active Python.
When you move from the default install to visible removal you reinstall with the extra, using the force flag so uv replaces the existing tool:
uv tool install --force "remove-ai-watermarks[visible]"The package requires Python 3.11 or newer and below 3.15, and the pyproject classifiers list 3.11 through 3.14. Further extras exist for narrower needs: heif, trustmark, migan, lama, diffusion and classify-onnx. Their exact dependency composition, Python compatibility and model requirements live in docs/installation.md under feature extras. The cost of the extras layout is that a single tool install on your PATH is shaped by whichever extra you installed last, so a second job later means another install.
metadata overwrites the original image unless you pass -o
The visible command writes to the path you give it with -o, which is the safe form to use:
remove-ai-watermarks visible image.png -o clean.pngThe metadata command is where the file-handling rules differ. Without -o on an image, it overwrites the source in place. With -o, the cleaned copy goes where you point it.
remove-ai-watermarks metadata image.png --remove -o clean.pngVideo behaves the opposite way. The video metadata command accepts MP4, MOV, M4V, WebM, MKV, AVI and FLV, supports a check mode and a remove mode, and when -o is omitted it writes a file named after the source with a _clean suffix and preserves the original.
remove-ai-watermarks video metadata input.mp4 --check
remove-ai-watermarks video metadata input.mp4 --remove -o clean.mp4The video command does not transcode video or audio streams, so it only rewrites metadata. Inspection of MP4 and MOV covers the native TC260 AIGC tag in moov.udta.meta.keys/ilst, including a moov box placed after the media payload. It also covers the QuickTime-form meta variants that Doubao's iOS export writes: a bare meta box as a direct moov child, and a keyless hdlr=mdir metadata list.
The reader consequence is concrete. A shell loop that mixes image and video runs will keep originals for video and destroy them for images unless every image call carries -o. Keep a copy of anything you cannot regenerate before the first run. Also note that metadata stripping alone removes only the C2PA manifest. If the image also carries a hidden pixel watermark linked to that manifest, the all command removes both, and metadata does not.
SynthID is read from signed provenance, and no local pixel detector exists
The package takes a specific position on SynthID. Signed provenance is the supported route, and identify reads it. There is no local SynthID pixel detector in the package. Research on a periodic lattice expert sits under scripts/synthid_runtime/ and in docs/synthid-detector-research.md, but that is research code, not a shipped command.
The nearby classifier does not fill the gap either. classify-source gives a lightweight, abstaining source-export and model-version signal, and its own description says it is not a SynthID detector. classify is explicitly not provenance.
For the reader this means identify reports what signed provenance says and nothing about what the pixels may hold. After running a removal, the package gives you no local way to confirm that a SynthID mark is gone from the pixels. For video, the route to SynthID removal is the video,diffusion extra, which the project describes as oracle-certified VAE regeneration, plus a video invisible command that applies a calibrated video-pixel SynthID-removal profile and lists a GPU as recommended.
Plan around that limit. If you need evidence of what a file claims about itself, identify is the tool. If you need a verdict on a hidden watermark from pixels alone, this project explicitly does not offer one.
Invisible image removal regenerates the picture and requires CUDA
The invisible command regenerates an image to disrupt invisible watermarks. It is the only command in the table that lists a GPU as required, and it needs the qwen-zimage extra with CUDA. The all command runs visible, invisible and metadata removal together, so it inherits the same hardware pull, with a GPU listed as recommended.
Regeneration changes the pixels, which is different from clearing a metadata field or inpainting a small sparkle. Anyone who needs the output to stay close to the original file should treat the invisible and all commands as a different kind of operation from visible and metadata, and compare the result against the source themselves, because the README gives no fidelity number for the regenerated output.
There is a narrower non-diffusion path for one format family. Microsoft Paint and Photos InvisMark declarations are routed automatically to pixel regeneration. A specialized Python API can inspect and disrupt the validated local Watermarker.dll payload without diffusion, and the pixels extra provides direct local Paint InvisMark disruption.
The hosted option is raiw.cc, which runs this library with the GPU included and nothing to install. Visible mark and metadata removal at Standard output up to 12 MP are free there. Original resolution above 12 MP and invisible watermark removal are paid. So the split between free and paid on the hosted service mirrors the split between CPU and GPU work locally.
The classifiers are opt-in and built to abstain
Two commands look at pixels rather than provenance, and neither runs unless you ask. classify decides whether a photograph is AI or camera-made, with an optional provider guess. identify never starts it. You install the extra and call it directly:
uv tool install --force "remove-ai-watermarks[classify]"
remove-ai-watermarks classify image.pngThe second command, classify-source, gives a 10-provider source-export and model-version signal intended for use after metadata removal. It abstains rather than guessing when it lacks a basis. It has both a CLI form and a Python API:
remove-ai-watermarks classify-source image.pngimport remove_ai_watermarks as raiw
result = raiw.classify_source("image.png")
print(result.label, result.version_subfamily, result.reason)The result carries a label, a version subfamily and a reason string, so a script can log why a call was made or why the classifier declined. It needs the source-classify extra. Guides are in docs/photo-classify.md and docs/source-classify.md.
One operational detail comes from the example environment file. Photo-classifier weights are fetched from Hugging Face unless RAIW_CLASSIFY_WEIGHTS points at a local directory. On a machine that cannot reach Hugging Face, set that variable to a directory that holds the weights, and read docs/photo-classify.md for what the directory must contain. HF_TOKEN is listed as optional and needed only for gated or private models.
The scope is your own content, and some jurisdictions forbid the tool outright
The project describes itself as a research project on AI provenance signals, for lawful use on content you own. It is provided as is, without warranty or liability, and each user answers for how they use it. It does not target stock agency previews or other watermarks that protect third party paid content. It also states that some jurisdictions, including China, restrict removing AI labels or supplying tools that do. Scope, safety and legal notes are in docs/legal-and-safety.md.
That statement changes who should run it. If you generate images with Gemini or another vendor and want to publish them without a sparkle or embedded generation fields, the intended use fits. If the file is someone else's licensed preview, the project says it is out of scope. If you operate in a place with labeling rules, the legal question comes before the technical one and is yours to settle.
The tool also has a content boundary. The command table, the description and the extras all cover images and video. None of the fourteen commands is a text tool, so searches for removing AI marks from text or Word documents will not be answered by this package. Likewise the visible video removal list is fixed to the registered marks named earlier, so an unregistered label on a clip is not covered by video visible.
Releases land every couple of days, and the PyPI name is pinned by downstream users
The last push was 2026-11-01 and the repository is not archived. Release v0.53.0 was published on 2026-11-01, v0.52.0 on 2026-10-30, and a tag named source-classify-freeze-2026-11-01, titled as the multi-provider source classifier v3 freeze, landed the same day as v0.53.0. pyproject.toml carries version 0.53.0 and the Development Status :: 4 - Beta classifier, so expect the command set to keep changing between versions.
A comment in pyproject.toml explains why the name will not change: the distribution name is the published PyPI package, and renaming it would strand every pip install and every pinned dependency, including the ComfyUI node package. The console script carries the same weight because users have it on PATH.
The license is Apache-2.0. For agent workflows the project ships a skill that installs with one command:
npx skills add wiltodelta/remove-ai-watermarksThe repository also holds plugin folders for Claude and Codex, a SECURITY.md, a uv.lock and a tests directory. The example environment file lists keys for OpenAI, Gemini and Azure Content Safety, but each is labeled as optional and tied to development scripts, such as generating provider samples or checking Microsoft Content Provenance, not to normal CLI use.
Given the release pace, pin the version in anything automated, and re-run your own check on a sample file after each upgrade.
Editorial conclusion
Use it if you own the images or video and need to clear Gemini-style visible marks or C2PA, EXIF and IPTC metadata without a GPU. Skip it for text, for other people's paid content, or if you expect a local SynthID detector. Check your jurisdiction first, and confirm whether -o is set before any metadata run on an image.
Frequently asked questions
How can I remove an AI watermark from an image with remove-ai-watermarks?
Install the visible extra and run remove-ai-watermarks visible image.png -o clean.png for known visible marks. For AI metadata, run remove-ai-watermarks metadata image.png --remove -o clean.png.
Is it possible to remove an AI watermark with remove-ai-watermarks?
Visible labels and metadata are removed without a GPU. Invisible pixel watermarks are handled by regeneration, which needs CUDA for the invisible command. The package has no local SynthID pixel detector to confirm the result.
Can remove-ai-watermarks remove AI watermarks from video?
Yes. It covers visible Sora, Veo, Seedance, Doubao, Dola, Hailuo AI and Kling AI marks, AI metadata, directory batches, and video SynthID removal through the video,diffusion extra.
Can I remove AI watermarks for free with remove-ai-watermarks?
The library is Apache-2.0 and installs with uv. On the hosted raiw.cc service, visible mark and metadata removal at Standard output up to 12 MP are free, while invisible removal and resolution above 12 MP are paid.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/wiltodelta-remove-ai-watermarks)Star history chart for your README
[](https://hysenlabs.com/projects/wiltodelta-remove-ai-watermarks)Was this useful?