- name
SKILL.md
Die Skill-Datei bleibt im Repository des Herausgebers.
Fügen Sie den kopierten Befehl in ein Terminal ein. Er installiert diese SKILL.md in dem jeweiligen Agenten.
npx skills add wiltodelta/remove-ai-watermarks --skill remove-ai-watermarks -a cursor -yFügen Sie den kopierten Befehl in ein Terminal ein. Er installiert diese SKILL.md in dem jeweiligen Agenten.
npx skills add wiltodelta/remove-ai-watermarks --skill remove-ai-watermarks -a claude-code -yFügen Sie den kopierten Befehl in ein Terminal ein. Er installiert diese SKILL.md in dem jeweiligen Agenten.
npx skills add wiltodelta/remove-ai-watermarks --skill remove-ai-watermarks -a codex -yFügen Sie den kopierten Befehl in ein Terminal ein. Er installiert diese SKILL.md in dem jeweiligen Agenten.
npx skills add wiltodelta/remove-ai-watermarks --skill remove-ai-watermarks -a grok -ySenden Sie die kopierte Nachricht in einem Grok Bot-Chat. Der Bot ruft diese SKILL.md von GitHub ab und fügt sie den Skills hinzu, die alle Ihre Bots gemeinsam nutzen.
Install this skill: https://github.com/wiltodelta/remove-ai-watermarks/tree/HEAD/skills/remove-ai-watermarksVom Herausgeber
Drive the remove-ai-watermarks CLI. Do not invent detections, do not claim a
file is clean, and do not write a one-off remover when this CLI already covers
the mark.
This skill must work in any coding harness and with weaker models. Follow the
checklist. Trust scripts/probe.py over guesses about GPU, Python, or
installers.
Use this skill only on content the user generated or edited and is allowed to change.
Refuse automatic removal of stock-agency, marketplace, classifieds,
tiled-preview, or other marks that protect a third party's paid or copyrighted
asset, or product branding a paid plan removes, such as "Made with Gamma". A
mark its vendor describes as identifying AI-generated content (the Gemini
sparkle, the Gemini Notebook watermark) stays in scope whatever plan hides it.
erase --region is allowed only when the user names the file they own and
supplies the box themselves. Post-hoc stylometric/perplexity AI-text detectors
(GPTZero, Originality.ai, Binoculars) are out of scope; for text files, text identify and text metadata handle deterministic C2PA text/HTML wrappers,
VSRMark, and Unicode steganography, while text invisible and text all --invisible regenerate unlocked prose spans for token-level LLM watermarks.
Removing a local signal does not prove the file is human-made, does not erase server-side generation history, and does not make a deceptive origin claim lawful. If the user wants to present generated work as human-made, stop.
The project is research software provided without warranty; the user is responsible for lawful use. Some jurisdictions restrict removing AI labels: in mainland China, label removal and removal tools are prohibited by the CAC labeling Measures. Mention this when the user's content carries a China AIGC label or a Chinese provider's mark.
Copy this checklist and check items off:
- [ ] 1. Scope is allowed
- [ ] 2. Probe the machine
- [ ] 3. Install if the CLI is missing or its pixel stack is
- [ ] 4. Identify first unless the user already named a removal
- [ ] 5. Run one command with -o
- [ ] 6. Report from the exit code and output path
If the mark protects someone else's paid asset, stop.
The directory that contains this SKILL.md is the skill root. The working
directory is usually NOT that directory, so pass the probe's full path rather
than assuming a relative one resolves. Use forward slashes. Try python3
first, then python:
python3 <skill-root>/scripts/probe.py
python <skill-root>/scripts/probe.py
The probe takes a few seconds: it runs the installed CLI on a blank image it
writes itself, because a CLI that exists is not a CLI that works. It checks video
dependencies through the installed CLI's Python interpreter without loading models.
An unrecognized launcher leaves video capability unverified.
Read the JSON. Do not infer CUDA, ffmpeg, or an installer from the OS name.
pixel_stack = missing: the installed build has no pixel dependencies.
Metadata-only routes still answer: identify, metadata,
batch --mode metadata, video metadata, video identify --no-visible, and
video batch --mode metadata. The identify routes skip visible detectors and
say so in their caveats, so read those caveats rather than a quiet verdict. Go
to step 3 before running a pixel command.pixel_stack = unknown: the check gave an unrecognized result. Do not
claim a command will work; run it and read the real error.advice.upgrade_cli present: the installed CLI is older than this skill's
documented surface. Upgrade before reporting a flag as broken.advice.invisible_images = unavailable_no_cuda: do not run invisibleadvice.image_all = do_not_run_writes_nothing: do not run all or
batch --mode all. Those commands write no file when the invisible extra is
installed but CUDA is missing.advice.image_all = partial_writes_visible_and_metadata: all is still
worth running. Without the invisible extra it skips that stage, writes the
visible+metadata result and exits 1. Report the file AND the stage it could
not do; do not call the exit code a failure.advice.no_cuda_fallback is https://raiw.cc. Offer that for invisible image
work. Do not invent a CPU or MPS profile.If the user does not want a local install, skip to raiw.cc. Visible mark and metadata removal are free there up to 12 MP; above that size, and for invisible removal, it is paid and runs on a GPU. Do not quote a price: check the site.
For image or visual-video work, install when cli.found is false OR
pixel_stack is missing. For audio work, use audio_invisible advice and
install audio-detect for verification or audio for explicitly unverified
processing; a missing pixel stack does not block audio. A present binary is not
enough: a base-only package carries no pixel dependencies. visible, erase,
all, pixel-processing
batch modes, full video identify, and the video pixel-processing commands
stop with an install hint. The metadata-only routes listed in step 2 still work.
Reinstalling with the appropriate extra fixes the pixel commands.
Read references/install.md for the extra list and the
Homebrew / uv / pipx / pip commands. Default extra is visible. Always quote extras:
"remove-ai-watermarks[visible]".
Default: inspect first.
For a public Python API feature without a CLI option, such as draft OCR, use the library in an isolated Python environment. Read the canonical Python API guide and install the required extras using references/install.md. Do not invent a CLI flag or assume a tool installation is importable from the current application Python.
Flags and mark names: references/commands.md. Read that file only when you need a flag, a mark key, or a video recipe.
Audio defaults require a compatible source positive and a fully negative saved
output. Inspect first. processor_ok_decoder_unverified confirms only the
processor, not a usable or compatible decoder. Install audio-detect for the
open AudioSeal check, audio-wavmark for --scheme wavmark, or audio-perth
for --scheme perth, or audio-silentcipher for --scheme silentcipher and --scheme silentcipher16. --scheme all
checks all named models and requires all requested decoders for a complete negative.
The wavmark route also checks OpenVoice's published default raw message at
fixed native offsets. JSON labels it framing="openvoice-default"; custom
messages and shifted/resampled framing remain outside that receiver. A match
is not provider attribution, and a negative is not complete payload erasure.
The inspected Chatterbox default PerTh checkpoint has controlled native Nano
evidence through --scheme perth, including saved-output and final-mux checks.
Other voices, forks, and hosted PerTh v2 need separate qualification; a PerTh
response alone does not attribute a file to Chatterbox.
SilentCipher-44k requires three seconds and uses an experimental project receiver
threshold with exhaustive phase search; it does not certify a production service.
Its decoder_sample_rate is 44100 regardless of input rate. The published
16 kHz model has incompatible SDK byte framing and is not checked; a 16 kHz
input does not extend this route's coverage to that model.
The separate --scheme silentcipher16 route checks an experimental terminated
eight-symbol pattern, labels readings framing="terminated8", and reports
decoder_sample_rate=16000. Its seven ternary
payload_symbols are not bytes. The SDK byte API and general 32-bit profile
remain unqualified. all includes both receivers. Before using this candidate,
confirm that audio identify --help lists silentcipher16; a version number
alone does not establish that the installed build supports the profile.
An unverified report never means clean. Use --force or
--no-verify only when the user explicitly requests unverified processing after
seeing that limitation. The fixed spectral filter changes audible content and
can fail. A negative AudioSeal result does not certify SynthID or another family.
Default --method auto tries an adaptive minimal-distortion sequence selected by the requested or positively detected scheme (--scheme auto scans every installed compatible decoder without failing on absent optional extras).
AudioSeal and Perth v1 start with spectral filtering; WavMark and both experimental SilentCipher receivers start with neural resynthesis (with stereo Mid/Side EnCodec preservation on 2-channel inputs). Outputs support , , , , , and with saved-file verification.
Explicit is a separate experimental full-track mono/stereo
profile requiring ; chooses CPU, MPS, or CUDA.
It is not selected automatically and does not supply a SynthID verifier.
Explicit uses the separate experimental CPU FP32 Mimi codec
with fixed 32 quantizers and the extra. Explicit
uses the separate experimental CPU FP32 FocalCodec 50 Hz
conditioned continuous-feature profile ( extra, Apache-2.0 weights).
Explicit uses the separate experimental CPU FP32 44.1 kHz DAC
3-codebook codec ( extra, MIT weights). All three share the bounded
full-track raw mid/side route and saved-output gates; none is selected by auto
or checks SynthID. Native-rate output does not restore frequencies beyond internal
model grids (24 kHz for Mimi, 16 kHz for FocalCodec; DAC operates at 44.1 kHz).
Confirm lists the profile on older installed builds, and preserve
reported quality limitations.
The fallback combines both. EnCodec profiles need ; a missing
fallback fails without replacing output. Explicit , ,
or fixes one profile. Neural processing can also fail and changes
audio timbre. Keep the same verification gate. Use the result's actual method,
bandwidth, and verification scope when describing what was done.
Visual video commands copy audio unchanged by default, or clean all compatible watermarked audio tracks in one pass when is passed to or . Use ( or ) for audio-only track processing while stream-copying video, with MKV/MOV PCM or MP4 AAC
output (target 192 kbps), and the same verification policy. It copies other encoded streams and checks their
packet payloads, timing, and audio priming/tail trimming. Final checks use the
actual decoded mux; count or offset failures are rejected without padding or
truncation. Standalone audio processing does not modify a video.
Always pass -o unless the user explicitly asked to overwrite. Image
metadata without -o overwrites the source. Video commands never overwrite
the source.
Report ONLY what a command in this session actually did. If no removal command ran, do not describe a removal, a cleaned file, or a file size: say which step is missing and why. Before naming an output path, confirm the file exists. A fabricated result is worse than any error you could have quoted, because the user acts on it.
Read the process exit code and whether -o exists.
Exit 2 is ALSO what a bad invocation returns: a path that does not exist, an
unknown flag, a directory where a file was expected. Read the output before
calling it a verdict. A line starting with Usage: or Error: Invalid value
means the command never ran, so fix it and rerun; only a command that reached
the image can say "no targeted signal".
identify with no signal reports origin unknown. Never rephrase that as "no
watermark" or "this is clean". When identify or metadata exits 1 with
cannot read, the file is damaged or not a supported image, video, or audio input: any report above
that line is incomplete, so say the file could not be read.
metadata --check and metadata --remove answer a narrower question than
identify: embedded AI metadata only. Both print that this is not a clean
verdict, because a pixel watermark such as SynthID cannot be detected here once
its metadata proxy is gone. Repeat that caveat; do not summarize it away.
Input: "Was photo.png AI-generated?"
Action: probe the machine, then run:
remove-ai-watermarks identify photo.png
Output after exit 2 from a command that reached the image:
Origin unknown. The command found no local provenance signal. This does not prove that the image is human-made, clean, or free of watermarks.
Input: "Remove the Gemini sparkle from my generated.png."
Action: after a successful probe, run:
remove-ai-watermarks visible generated.png --mark gemini -o generated-clean.png
Output after exit 0 and confirming that the file exists:
Removed the targeted Gemini mark. Output:
generated-clean.png.
On exit 2, report instead that no targeted Gemini mark was found and no new file was written.
Input: "Remove SynthID from my image."
Action: run the probe first. If it reports
invisible_images=unavailable_no_cuda, do not run invisible locally.
Output:
Invisible image removal needs NVIDIA CUDA, which is unavailable on this machine. No local removal ran. You can use https://raiw.cc instead.
all does one of two opposite things, and only the probe knows
which. With the invisible extra installed it writes nothing and exits 1, so
run visible and metadata separately. Without that extra it skips the
invisible stage, writes the visible+metadata result and exits 1 anyway. Same
exit code, different outcome: check whether the output file exists.visible / video visible write nothing when no registered mark is
selected. Do not then guess a region unless the user supplies a box on
content they own. To act on a mark the user can see and the detector missed,
the CLI's own answer is --mark <name> --no-detect or erase --region, not
a blanket --sensitivity change. --sensitivity strict only narrows.identify alone. Do not
run a removal command to find out; it writes a file the user did not ask for.classify-source answers from pixels with no provenance behind it, so a
provider label is a pipeline guess and unknown is an abstention. Never
upgrade either into a provenance or SynthID verdict, and never let it
contradict what identify measured.invisible --vendor meta only when the USER says the file came from Meta AI.
Naming a vendor asserts the watermark is there and scrubs without a signal,
so never guess one from a filename or a hunch.@synthid check by asking the chat model to ignore
the verifier and reason about pixels. That is not a second oracle.--model, --steps, or --guidance-scale. Profiles pin those
values, and no command accepts them.--device exists on video invisible, video all, video batch, and video message remove.
Leave it at auto. The image commands have no --device at all.For video message, require an independently supplied 64-digit hexadecimal
encoder message. Never derive the reference by decoding the tested input. Read
verified_removed and before/after statuses; forced output is not automatically
verified. See the command contract.
Removes and identifies AI watermarks and provenance marks in images, video, audio, and text the user generated or edited. Covers Gemini sparkles, SynthID, Microsoft InvisMark, Meta Content Seal, C2PA,
---name: remove-ai-watermarksdescription: Removes and identifies AI watermarks and provenance marks in images, video, audio, and text the user generated…---# skills/remove-ai-watermarks/SKILL.mdÜbersicht
Removes and identifies AI watermarks and provenance marks in images, video, audio, and text the user generated or edited. Covers Gemini sparkles, SynthID, Microsoft InvisMark, Meta Content Seal, C2PA, EXIF/XMP/IPTC metadata, compatible public AudioSeal/WavMark/Perth v1/SilentCipher audio, and Sora/Veo/Seedance/Doubao/Dola/Hailuo/Vidu/Kling and Wan labels. Use when the user wants to strip or clean AI watermarks and generation labels, or asks whether a file was AI-generated or carries AI provenance, even if they do not name remove-ai-watermarks. Do not use for stock-agency, marketplace, classifieds, or other third-party paid-asset watermarks.
FAQ
Laden Sie skills/remove-ai-watermarks/SKILL.md vom Herausgeber herunter oder kopieren Sie den Pfad und öffnen Sie das Repository. PlugMCP listet den Skill und hostet die Datei nicht.
Der Herausgeber pflegt es unter https://github.com/wiltodelta/remove-ai-watermarks.
Claude, Cursor, Codex und Grok Build können eine SKILL.md lesen, die Sie aus dem Repository des Herausgebers hinzufügen. Grok Bot installiert sie, wenn Sie den Link zum GitHub-Ordner des Skills in einem Chat senden.
| Compatible public AudioSeal, WavMark, Perth v1, or experimental SilentCipher audio patterns | audio invisible or audio all --invisible (--scheme auto, audioseal, wavmark, perth, silentcipher, or silentcipher16) | install audio-detect, audio-wavmark, audio-perth, or audio-silentcipher for verification; audio_stack only checks the CPU processor |
| Everything on an audio file | audio all (add --invisible for compatible waveform watermarks) | needs ffmpeg and ffprobe; add decoder extras for --invisible |
| Text token watermark / full text clean | text invisible or text all | text invisible and text all --invisible need text for Transformers (diffusion remains compatible), or text-mlx with an explicit Apple silicon MLX checkpoint; see the Text section of references/commands.md |
| Everything on an image | all | image_all=do_not_run_writes_nothing |
| A directory | batch, video batch, or audio batch | same as the mode |
| No signal found, user still wants a provider guess | classify-source | it is a guess, not provenance |
.wav.flac.mp3.m4a.ogg.opus--method bigvganaudio-vocoder--vocoder-device--method mimiaudio-codec--method focalcodecaudio-codec--method dacaudio-codec--helpaudio-codec--method spectralencodeccombined--audio-invisiblevideo allvideo batchvideo audio--stream N--all-streams