Skip to content
Back to search

remove-ai-watermarks

MarketplaceWorks with Claude CodeGitHubIncomplete

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today

Check it deeper on your machine

It has changed since its scan on 17 Nov 2026.

As a Claude Code plugin, install it and then ask Claude Code:

/plugin marketplace add keithmackay/mackayi/plugin install scanskillsafety@mackayiDo a deep scan of https://github.com/wiltodelta/remove-ai-watermarks with scanskillsafety

Or from a terminal, clone it and run it:

git clone https://github.com/keithmackay/scanskillsafety-skill ~/.claude/skills/scanskillsafetynode ~/.claude/skills/scanskillsafety/cli.cjs --deep 'https://github.com/wiltodelta/remove-ai-watermarks'

It checks the exact commit: every file, its packages against OSV (api.osv.dev), how it compares with this page's scan (findsafeskills.com receives the repository's address), and, if you agree, Claude reading every file with your own Claude subscription. It never runs the code.

Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/wiltodelta--remove-ai-watermarks/badge.svg)](/listings/wiltodelta--remove-ai-watermarks)

HTML

<a href="/listings/wiltodelta--remove-ai-watermarks"><img src="/api/listings/wiltodelta--remove-ai-watermarks/badge.svg" alt="findsafeskills safety rating" /></a>

Install

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

| Need | Install | | --- | --- | | Metadata inspection and stripping | `remove-ai-watermarks` | | 10-provider source-export & model-version classification | `remove-ai-watermarks[source-classify]` | | Visible detection and removal | `remove-ai-watermarks[visible]` | | Visible video processing | `remove-ai-watermarks[video]` | | Public known-message VideoSeal/PixelSeal checks | `remove-ai-watermarks[video-detect]`, add `diffusion` for VAE | | Video SynthID removal | `remove-ai-watermarks[video,diffusion]` | | Audio sample processing and spectral disruption | `remove-ai-watermarks[audio]` | | Compatible public audio watermark detection (AudioSeal) | `remove-ai-watermarks[audio-detect]` | | Neural audio codec resynthesis (EnCodec / FocalCodec / DAC) | `remove-ai-watermarks[audio-codec]` | | Torch-free DWT-DCT detection | `remove-ai-watermarks[detect]` | | Direct local Paint InvisMark disruption | `remove-ai-watermarks[pixels]` | | Invisible image removal (needs CUDA) | `remove-ai-watermarks[qwen-zimage]` | | Production removal and default classifier backends | `remove-ai-watermarks[all]` | | Optional OCR proposals for operator verification | `remove-ai-watermarks[text-draft]` |

Lower-level and specialized extras include `pixels`, `heif`, `trustmark`, `migan`, `lama`, `diffusion`, `source-classify`, `audio-vocoder`, `audio-wavmark`, `audio-perth`, and `audio-silentcipher`. The [installation guide](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/installation.md#feature-extras) documents their exact dependency composition, Python compatibility, and model requirements.

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

.claude-plugin/marketplace.json · Manifest · 19 lines
  1. {
  2. "name": "remove-ai-watermarks",
  3. "owner": {
  4. "name": "wiltodelta",
  5. "url": "https://github.com/wiltodelta"
  6. },
  7. "metadata": {
  8. "description": "Agent skill and Claude Code plugin for the remove-ai-watermarks CLI.",
  9. "version": "1.0.21"
  10. },
  11. "plugins": [
  12. {
  13. "name": "remove-ai-watermarks",
  14. "description": "Identify and remove AI provenance marks from images, video, audio, and text the user generated or edited. Use when the user asks to strip Gemini sparkles, SynthID, Microsoft InvisMark, Meta Content Seal, C2PA, EXIF/XMP AI metadata, AudioSeal/WavMark/Perth/SilentCipher audio marks, Sora/Veo/Seedance/Doubao/Dola/Hailuo/Vidu/Kling labels, or other AI watermarks, or asks whether a file was AI-generated. Do not use for stock-agency, marketplace, classifieds, or other third-party paid-asset watermarks.",
  15. "source": "./skills",
  16. "strict": false
  17. }
  18. ]
  19. }
README.md · README · 656 lines
  1. # Remove AI Watermarks
  2. Remove AI provenance marks from images, video, and audio you generated yourself:
  3. - known visible labels such as the Google Gemini sparkle watermark and vendor text marks;
  4. - invisible pixel watermarks through direct local-format disruption or diffusion regeneration;
  5. - compatible public audio watermarks, with source and saved-output decoder checks;
  6. - C2PA, EXIF, XMP, IPTC, and related AI metadata.
  7. | Before | After |
  8. | --- | --- |
  9. | ![Image with a visible Gemini watermark](https://raw.githubusercontent.com/wiltodelta/remove-ai-watermarks/main/demo_banana_before.png) | ![Image after visible watermark removal](https://raw.githubusercontent.com/wiltodelta/remove-ai-watermarks/main/demo_banana_after.png) |
  10. ```bash
  11. uv tool install "remove-ai-watermarks[visible]" # CLI for visible marks and metadata
  12. npx skills add wiltodelta/remove-ai-watermarks # agent skill for Claude Code, Codex, Cursor, and other agents
  13. ```
  14. Video support covers provenance identification, complete visible-plus-metadata
  15. cleaning, directory batches, visible Sora, Veo, Seedance, Doubao, Dola,
  16. Hailuo AI, and Kling AI mark removal, and oracle-certified VAE regeneration for video SynthID
  17. removal.
  18. > [raiw.cc](https://raiw.cc) runs this library as a hosted service, with the GPU
  19. > included and nothing to install. Visible mark and metadata removal at Standard
  20. > output up to 12 MP are free there; original resolution above 12 MP and invisible
  21. > watermark removal are paid.
  22. [![PyPI](https://img.shields.io/pypi/v/remove-ai-watermarks?logo=pypi&logoColor=white)](https://pypi.org/project/remove-ai-watermarks/)
  23. [![Python](https://img.shields.io/pypi/pyversions/remove-ai-watermarks?logo=python&logoColor=white)](https://pypi.org/project/remove-ai-watermarks/)
  24. [![Downloads](https://static.pepy.tech/badge/remove-ai-watermarks/month)](https://pepy.tech/project/remove-ai-watermarks)
  25. [![License](https://img.shields.io/pypi/l/remove-ai-watermarks?color=blue)](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/LICENSE)
  26. [![Tests](https://github.com/wiltodelta/remove-ai-watermarks/actions/workflows/test.yml/badge.svg)](https://github.com/wiltodelta/remove-ai-watermarks/actions/workflows/test.yml)
  27. [![Sponsor](https://img.shields.io/badge/Sponsor-GitHub-db61a2?logo=githubsponsors&logoColor=white)](https://github.com/sponsors/wiltodelta)
  28. [![skills.sh](https://skills.sh/b/wiltodelta/remove-ai-watermarks)](https://skills.sh/wiltodelta/remove-ai-watermarks)
  29. [![Glama MCP server](https://glama.ai/mcp/servers/wiltodelta/remove-ai-watermarks/badges/score.svg)](https://glama.ai/mcp/servers/wiltodelta/remove-ai-watermarks)
  30. > This is a research project on AI provenance signals, for lawful use on content
  31. > you own. It is provided as is, without warranty or liability, and each user is
  32. > responsible for how they use it. It does not target stock agency previews or
  33. > other watermarks that protect third party paid content. Some jurisdictions,
  34. > including China, restrict removing AI labels or supplying tools that do. See
  35. > [scope, safety, and legal notes](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/legal-and-safety.md).
  36. ## Choose what you want to do
  37. | Goal | Command | GPU |
  38. | --- | --- | --- |
  39. | Find provenance signals and watermarks | `identify` | No |
  40. | Classify likely AI provider and model version from pixels (opt-in, not provenance) | `classify-source` | No |
  41. | Remove known visible AI marks | `visible` | No |
  42. | Erase a region you select | `erase` | No |
  43. | Strip AI metadata | `metadata` | No |
  44. | Identify supported video provenance | `video identify` | No |
  45. | Inspect audio metadata and compatible public-model evidence | `audio identify` (`--scheme auto`, `audioseal`, `wavmark`, `perth`, `silentcipher`, `silentcipher16`, or `all`) | No |
  46. | Strip AI metadata from audio | `audio metadata` | No |
  47. | Process a detected compatible audio watermark and verify the saved file (`.wav`, `.flac`, `.mp3`, `.m4a`, `.ogg`, `.opus`) | `audio invisible` | No |
  48. | Strip AI metadata and optionally clean compatible audio watermarks | `audio all` / `audio batch` | No |
  49. | Process one or all audio tracks (`--all-streams`) while copying encoded video | `video audio` | No |
  50. | Remove visible marks, AI metadata, and opt-in video/audio invisible marks | `video all` | Depends on `--invisible` |
  51. | Strip AI metadata from video | `video metadata` | No |
  52. | Remove a registered visible AI mark from video | `video visible` | No |
  53. | Process a directory of videos | `video batch` | Depends on mode |
  54. | Inspect and disrupt a known public VideoSeal/PixelSeal message (experimental) | `video message` | Optional |
  55. | Regenerate video pixels (`--profile auto`, calibrated SynthID, Seedance, or native CUDA `Wan 2.1`) | `video invisible` | Recommended |
  56. | Identify C2PA text/HTML wrappers, `VSRMark`, and Unicode steganography | `text identify` | No |
  57. | Strip C2PA text/HTML wrappers, `VSRMark`, and Unicode steganography | `text metadata` | No |
  58. | Regenerate unlocked prose for best-effort token-watermark removal | `text invisible` | Recommended |
  59. | Strip text provenance and optionally regenerate prose spans | `text all` | Depends on `--invisible` |
  60. | Regenerate an image to disrupt invisible watermarks | `invisible` | Required (CUDA) |
  61. | Run visible, invisible, and metadata removal | `all` | Recommended |
  62. | Process a directory | `batch` | Depends on mode |
  63. Install `remove-ai-watermarks[text]` for the default local Transformers text
  64. rewriter and its packaged language checker. Local text regeneration also supports
  65. optional Apple silicon MLX models through
  66. `remove-ai-watermarks[text-mlx]` and
  67. `--rewriter-backend mlx --rewriter-model /path/to/checkpoint`.
  68. Both backends share the same acceptance guards; model output still needs factual
  69. review. See the [text CLI](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/cli.md#identify-and-clean-text).
  70. Experimental public VideoSeal and PixelSeal known-message video checks and processing
  71. use [`video message`](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/cli.md#public-model-video-messages-experimental).
  72. They require an independently known encoder message and do not certify arbitrary
  73. unknown marks or provider deployments.
  74. Microsoft Paint and Photos InvisMark declarations are routed automatically to
  75. pixel regeneration. The `all` command removes both the hidden pixel watermark and
  76. its linked C2PA manifest; metadata stripping alone removes only the manifest. A
  77. specialized Python API can inspect and disrupt the validated local
  78. `Watermarker.dll` payload without diffusion.
  79. ## Installation modes
  80. | Need | Install |
  81. | --- | --- |
  82. | Metadata inspection and stripping | `remove-ai-watermarks` |
  83. | 10-provider source-export & model-version classification | `remove-ai-watermarks[source-classify]` |
  84. | Visible detection and removal | `remove-ai-watermarks[visible]` |
  85. | Visible video processing | `remove-ai-watermarks[video]` |
  86. | Public known-message VideoSeal/PixelSeal checks | `remove-ai-watermarks[video-detect]`, add `diffusion` for VAE |
  87. | Video SynthID removal | `remove-ai-watermarks[video,diffusion]` |
  88. | Audio sample processing and spectral disruption | `remove-ai-watermarks[audio]` |
  89. | Compatible public audio watermark detection (AudioSeal) | `remove-ai-watermarks[audio-detect]` |
  90. | Neural audio codec resynthesis (EnCodec / FocalCodec / DAC) | `remove-ai-watermarks[audio-codec]` |
  91. | Torch-free DWT-DCT detection | `remove-ai-watermarks[detect]` |
  92. | Direct local Paint InvisMark disruption | `remove-ai-watermarks[pixels]` |
  93. | Invisible image removal (needs CUDA) | `remove-ai-watermarks[qwen-zimage]` |
  94. | Production removal and default classifier backends | `remove-ai-watermarks[all]` |
  95. | Optional OCR proposals for operator verification | `remove-ai-watermarks[text-draft]` |
  96. Lower-level and specialized extras include `pixels`, `heif`, `trustmark`,
  97. `migan`, `lama`, `diffusion`, `source-classify`, `audio-vocoder`,
  98. `audio-wavmark`, `audio-perth`, and `audio-silentcipher`. The
  99. [installation guide](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/installation.md#feature-extras) documents their exact
  100. dependency composition, Python compatibility, and model requirements.
  101. ## Quick start
  102. Install the metadata-focused default CLI:
  103. ```bash
  104. uv tool install remove-ai-watermarks
  105. ```
  106. Inspect an image:
  107. ```bash
  108. remove-ai-watermarks identify image.png
  109. ```
  110. For an abstaining, CPU-only 10-provider source-export and model-version
  111. classifier from pixels after metadata removal (`classify_source` and CLI
  112. `classify-source`), install `source-classify`. `identify` never runs it, and it
  113. is not a SynthID detector:
  114. ```bash
  115. uv tool install --force "remove-ai-watermarks[source-classify]"
  116. remove-ai-watermarks classify-source image.png
  117. ```
  118. ```python
  119. import remove_ai_watermarks as raiw
  120. result = raiw.classify_source("image.png")
  121. print(result.label, result.version_subfamily, result.synthetic_detected, result.confidence)
  122. ```
  123. Guide: [source-pipeline classification](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/source-classify.md). Model card: [`wiltodelta/raiw-source-classify`](https://huggingface.co/wiltodelta/raiw-source-classify).
  124. Signed provenance is the supported route for SynthID and `identify` reads it.
  125. There is no local SynthID pixel detector in the package. Research on a
  126. periodic lattice expert is in `scripts/synthid_runtime/` and
  127. [synthid-detector-research.md](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/synthid-detector-research.md).
  128. For visible watermark removal, install the pixel dependencies:
  129. ```bash
  130. uv tool install --force "remove-ai-watermarks[visible]"
  131. ```
  132. Then remove a known visible mark and AI metadata:
  133. ```bash
  134. remove-ai-watermarks visible image.png -o clean.png
  135. ```
  136. Strip metadata without running visible inpainting or diffusion:
  137. ```bash
  138. remove-ai-watermarks metadata image.png --remove -o clean.png
  139. ```
  140. Without `-o` this command overwrites the source in place.
  141. Inspect or remove AI metadata from an MP4, MOV, M4V, WebM, MKV, AVI, or FLV
  142. file:
  143. ```bash
  144. remove-ai-watermarks video metadata input.mp4 --check
  145. remove-ai-watermarks video metadata input.mp4 --remove -o clean.mp4
  146. ```
  147. The `video metadata` command does not transcode video or audio streams. Unlike
  148. the image command above, when `-o` is omitted it writes `<source>_clean` and
  149. preserves the original. MP4 and MOV
  150. inspection includes the native TC260 `AIGC` tag in
  151. `moov.udta.meta.keys/ilst`, including a `moov` placed after the media payload,
  152. plus the QuickTime-form `meta` variants Doubao's iOS export writes (a bare
  153. `meta` box as a direct `moov` child, and a keyless `hdlr=mdir` metadata list).
  154. Keyed `workflow` and `prompt` entries in the same metadata-list structure are
  155. also detected and removed, including ComfyUI exports.
  156. MKV and WebM inspection reads the normative
  157. `Segment.Tags.Tag.SimpleTag` placement. AVI uses `LIST/INFO/AIGC`, while FLV
  158. uses `script.onMetaData.AIGC`. The non-ISOBMFF formats are remuxed with stream
  159. copy for removal.
  160. Use the product-oriented video path to identify or clean a file:
  161. ```bash
  162. uv tool install --force "remove-ai-watermarks[video]"
  163. remove-ai-watermarks video identify input.mp4
  164. remove-ai-watermarks video all input.mp4 -o clean.mp4
  165. ```
  166. `video identify` also names invisible watermark hints from metadata and provider
  167. policy. These are separate from pixel detection and per-file removal verification.
  168. `video all` removes a stable registered visible mark when present and always
  169. strips verified AI metadata. If neither signal is found, it still writes a
  170. same-container passthrough, so application callers get one predictable output
  171. contract. Proprietary invisible-video removal is excluded by default:
  172. `--invisible` opts into the lossy, oracle-certified video SynthID profile (with motion-compensated temporal stabilization enabled by default and `--keep-fps` to retain native frame rate), and `--audio-invisible` inspects and cleans compatible watermarked audio tracks in the same atomic pipeline.
  173. Process a directory with the same contract:
  174. ```bash
  175. remove-ai-watermarks video batch ./videos --mode all
  176. ```
  177. Remove a supported visible video mark:
  178. ```bash
  179. remove-ai-watermarks video visible input.mp4 -o clean.mp4
  180. remove-ai-watermarks video visible veo.mp4 --mark veo -o veo_clean.mp4
  181. remove-ai-watermarks video visible seedance.mp4 --mark seedance -o seedance_clean.mp4
  182. remove-ai-watermarks video visible dola.mp4 --mark dola -o dola_clean.mp4
  183. remove-ai-watermarks video visible hailuo.mp4 --mark hailuo -o hailuo_clean.mp4
  184. remove-ai-watermarks video visible kling.mp4 --mark kling -o kling_clean.mp4
  185. ```
  186. This path scans the complete sequence before changing pixels. It accepts only a
  187. mark that repeats at a stable position across adjacent frames, then reuses the
  188. same OpenCV, MI-GAN, or LaMa fill backends as image removal. Audio is copied
  189. without re-encoding and is allowed to reach its natural end; the video stream
  190. is transcoded because its pixels change. By default, a guarded optical-flow
  191. pass motion-aligns the preceding accepted fill and blends it only when the
  192. nearby source context agrees; use `--no-temporal-consistency` to disable it.
  193. The encoder preserves supported
  194. 8-bit source chroma sampling, color tags, and MP4/MOV track timescale instead
  195. of relying on ffmpeg's implicit raw-BGR defaults. Variable frame intervals are
  196. preserved through a timestamped in-memory NUT bridge instead of being flattened
  197. to the average frame rate. Non-zero source start timestamps are retained
  198. together with the copied audio offset. The default `--mark auto`
  199. scans all providers in one decode pass and selects the first stable match in
  200. the specificity order shown below. Valid Google AI-video provenance breaks a
  201. Sora/Veo cross-match in favor of Veo when the diamond is nested in the Sora
  202. region or Veo persists through the complete clip while Sora does not.
  203. Pass an explicit mark to restrict detection to one provider.
  204. Sora covers the moving Sora 2 mascot, with or without the wordmark; mascot-only
  205. matches also require the complete two-eye shape. Veo covers both the current
  206. four-point diamond and the legacy `Veo` text. Seedance covers the fixed boxed
  207. `AI` label, Dola covers the fixed `Dola AI` text, Hailuo AI covers the composite
  208. `MINIMAX | hailuo AI` label, and Kling AI covers the bottom-right `KLING AI` or
  209. `KlingAI` label with its version suffix. A completed encode is published
  210. atomically. No output is written when no stable mark is found.
  211. HDR, PQ/HLG, and greater-than-8-bit inputs are rejected before encoding rather
  212. than silently reduced through OpenCV's 8-bit BGR boundary.
  213. Apply the calibrated video-pixel SynthID-removal profile:
  214. ```bash
  215. uv tool install --force "remove-ai-watermarks[video,diffusion]"
  216. remove-ai-watermarks video invisible input.mp4 -o clean.mp4
  217. ```
  218. This path regenerates the complete sequence with one latent-noise field shared
  219. across time, copies complete audio, strips source metadata, and publishes the
  220. completed encode atomically. The default `noise_std=0.15` profile passed both
  221. the two-carrier calibration and a complete public eight-second Veo oracle
  222. check. Google does not publish a local decoder, so a fresh provider check
  223. remains useful for unusually important files or after provider changes, but it
  224. is not a product result state. The CLI therefore describes video-pixel
  225. regeneration instead of claiming a per-file SynthID verdict, and reports the
  226. copied audio watermark as `UNVERIFIED`.
  227. For invisible watermark removal through the CLI or high-level API, install the
  228. `qwen-zimage` extra. **An NVIDIA GPU is required**: all diffusion profiles are
  229. CUDA-only, and there is no CPU or MPS fallback. The specialized local Paint API
  230. uses the `pixels` extra without CUDA.
  231. ```bash
  232. uv tool install --force "remove-ai-watermarks[qwen-zimage]"
  233. remove-ai-watermarks invisible image.png -o clean.png
  234. ```
  235. If the local detectors cannot confirm an invisible watermark but you know the
  236. image came from an AI generator, add `--force`:
  237. ```bash
  238. remove-ai-watermarks invisible image.png -o clean.png --force
  239. ```
  240. Typography-heavy images can opt into the experimental verified-text post-pass.
  241. It accepts manually reviewed strings and line boxes or an operator-verified
  242. geometry-only manifest; it never treats raw OCR output as ground truth or runs
  243. automatically:
  244. ```bash
  245. uv tool install --force "remove-ai-watermarks[text-restoration]"
  246. remove-ai-watermarks invisible image.png -o clean.png \
  247. --text-manifest verified-lines.json --force
  248. ```
  249. See the [CLI guide](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/cli.md#restore-operator-verified-text) for the manifest
  250. schema, compatibility restrictions, and oracle caveats.
  251. See the [installation guide](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/installation.md) for Homebrew, uv, optional
  252. features, and development setup.
  253. ## Examples
  254. ### Visible Gemini mark
  255. The `after` raster at the top of this page is generated from the tracked `before`
  256. raster by the public path:
  257. ```bash
  258. uv run remove-ai-watermarks visible demo_banana_before.png \
  259. --backend cv2 -o demo_banana_after.png
  260. ```
  261. ### High quality invisible removal
  262. `qwen-zimage` is the default profile: a Qwen-Image-2512 Lightning pass under Canny
  263. ControlNet, followed by SAM-masked Z-Image repair of any detected face. The
  264. alternative, `sdxl-zimage`, swaps the global stage for SDXL and keeps the same face
  265. stage. `chroma-zimage` uses the Apache-2.0 Chroma1 global pass with its own flat
  266. vendor floors; see `docs/chroma1-engine-research.md` for the calibration. An opt-in
  267. profile, `qwen21-zimage`, swaps the global stage for Qwen-Image-2.1 (with
  268. `madebyollin/texture-fix-vae-for-qwen-image-2.1` and `Qwen-Image-2.1-Fun-Controlnet-Union`
  269. Canny); because `Qwen/Qwen-Image-2.1` is released under the Qwen Research License,
  270. `qwen21-zimage` is for non-commercial research, evaluation, and personal use only
  271. (not for commercial use). `--pipeline auto` picks sdxl-zimage for Google,
  272. chroma-zimage for Microsoft, and qwen-zimage otherwise, keeping `auto` commercial-safe.
  273. All are CUDA only.
  274. Since 0.57.0, `qwen21-zimage` includes bounded geometric alignment by default after the shared
  275. Qwen-2.1 global and face stages, accepting a result only when both source MSE and local LPIPS improve.
  276. It retains the Qwen Research License restrictions. Qualified text geometry can
  277. be supplied through the Python API; the CLI does not run automatic OCR. See
  278. [the fidelity profile](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/cli.md#fidelity-profile) for limits and fallback behavior.
  279. ```bash
  280. uv tool install --force "remove-ai-watermarks[qwen-zimage]"
  281. remove-ai-watermarks invisible image.png -o clean.png --force
  282. ```
  283. | OpenAI example before | OpenAI example after |
  284. | --- | --- |
  285. | [![OpenAI portrait grid before qwen-zimage](https://raw.githubusercontent.com/wiltodelta/remove-ai-watermarks/main/data/synthid/originals/ChatGPT%20Image%20May%2030,%202026,%2010_31_08%20AM.png)](https://raw.githubusercontent.com/wiltodelta/remove-ai-watermarks/main/data/synthid/originals/ChatGPT%20Image%20May%2030,%202026,%2010_31_08%20AM.png) | [![OpenAI portrait grid after qwen-zimage](https://raw.githubusercontent.com/wiltodelta/remove-ai-watermarks/main/docs/images/qwen-zimage/ChatGPT/ChatGPT%20Image%20May%2030,%202026,%2010_31_08%20AM_full_clean.png)](https://raw.githubusercontent.com/wiltodelta/remove-ai-watermarks/main/docs/images/qwen-zimage/ChatGPT/ChatGPT%20Image%20May%2030,%202026,%2010_31_08%20AM_full_clean.png) |
  286. | Gemini example before | Gemini example after |
  287. | --- | --- |
  288. | [![Gemini sign before qwen-zimage](https://raw.githubusercontent.com/wiltodelta/remove-ai-watermarks/main/data/synthid/originals/Gemini_Generated_Image_633uuy633uuy633u.png)](https://raw.githubusercontent.com/wiltodelta/remove-ai-watermarks/main/data/synthid/originals/Gemini_Generated_Image_633uuy633uuy633u.png) | [![Gemini sign after qwen-zimage](https://raw.githubusercontent.com/wiltodelta/remove-ai-watermarks/main/docs/images/qwen-zimage/Gemini/Gemini_Generated_Image_633uuy633uuy633u_full_clean.png)](https://raw.githubusercontent.com/wiltodelta/remove-ai-watermarks/main/docs/images/qwen-zimage/Gemini/Gemini_Generated_Image_633uuy633uuy633u_full_clean.png) |
  289. These exact output files were checked with the matching provider verifiers. That
  290. result applies to these files, not to every seed, image, or future watermark
  291. version.
  292. ## Common recipes
  293. ### Remove every detected visible mark
  294. ```bash
  295. remove-ai-watermarks visible image.png -o clean.png
  296. ```
  297. The default `--mark auto` checks all registered visible marks and removes every
  298. match. If the mark is visible to you but the detector misses it, select its
  299. region explicitly:
  300. ```bash
  301. remove-ai-watermarks erase image.png \
  302. --region 1640,1930,400,100 \
  303. -o clean.png
  304. ```
  305. `--region` uses `x,y,width,height` and may be repeated.
  306. ### Use a learned fill backend
  307. The `visible` extra uses OpenCV inpainting when no learned backend is installed.
  308. For more difficult backgrounds, the learned-backend extras include the same
  309. pixel dependencies automatically:
  310. ```bash
  311. uv tool install --force "remove-ai-watermarks[migan]"
  312. remove-ai-watermarks visible image.png -o clean.png --backend migan
  313. ```
  314. ```bash
  315. uv tool install --force "remove-ai-watermarks[lama]"
  316. remove-ai-watermarks visible image.png -o clean.png --backend lama
  317. ```
  318. ### Reduce CUDA memory use
  319. ```bash
  320. remove-ai-watermarks invisible image.png -o clean.png \
  321. --cpu-offload --force
  322. ```
  323. CPU offload lowers CUDA memory pressure by moving model components between CPU
  324. and GPU, at the cost of speed.
  325. ### Process a directory
  326. ```bash
  327. remove-ai-watermarks batch ./images --mode visible
  328. remove-ai-watermarks batch ./images --mode all
  329. ```
  330. ## What the tool can recognize
  331. Visible mark support includes:
  332. - Google Gemini and Nano Banana visible sparkle watermark;
  333. - Doubao, the Jimeng wordmark and top-left `AI生成` layouts, Qwen, Kling AI,
  334. Yuanbao, Dola, WorkBuddy, Baidu, LiblibAI's bottom-center wordmark and compact
  335. top-left pill, and RunningHub labels;
  336. - calibrated Microsoft top-right white and beige AI-badge layouts;
  337. - calibrated Italian and Korean Samsung Galaxy AI disclosures;
  338. - a generic brand-less bottom-right `AI生成` fallback for OS/gallery-level
  339. AI-edit stamps that carry no vendor wordmark of their own (vivo and Xiaomi
  340. Gallery confirmed); synthetic-only calibration, weaker evidence than the
  341. vendor-tuned detectors above;
  342. - the OpenArt wordmark, placed at the frame center rather than a corner;
  343. uncalibrated (one confirmed real case, no captured corpus).
  344. Metadata and provenance inspection covers C2PA, EXIF, XMP, IPTC, common
  345. generator parameters in image and video containers, China TC260 AIGC labels,
  346. and several vendor specific signals. Optional decoders add support for open
  347. DWT-DCT watermarks and Adobe TrustMark.
  348. The exact support matrix, including important locale and detector limits, lives
  349. in [supported signals](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/supported-signals.md). The
  350. [synthetic example gallery](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/data/fixtures/visible/README.md) carries one canonical
  351. sample for every registered image and video mark.
  352. ## How it works
  353. Visible removal follows three steps:
  354. 1. Detect a registered mark in its expected area.
  355. 2. Build a mask around the mark.
  356. 3. Fill only the masked region with OpenCV, MI-GAN, or LaMa.
  357. Metadata removal uses format aware stripping. JPEG metadata removal preserves
  358. the encoded image scan instead of recompressing it. Native MP4/MOV TC260 values
  359. and keyed generation parameters are blanked without changing box sizes or
  360. media offsets. Other supported containers use their corresponding metadata
  361. path.
  362. Invisible removal through the CLI and high-level API regenerates the image through
  363. a diffusion pipeline. Specialized callers can instead decode and disrupt the
  364. positively identified local Microsoft Paint format with much smaller pixel changes.
  365. Neither path can guarantee that every proprietary verifier will reject every output.
  366. See [supported signals](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/supported-signals.md) and
  367. [known limitations](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/known-limitations.md) for the full technical boundary.
  368. ## Python API
  369. The visible-removal API requires `remove-ai-watermarks[visible]`.
  370. ```python
  371. import remove_ai_watermarks as raiw
  372. result, removed = raiw.remove_visible("watermarked.png", "clean.png")
  373. print(removed)
  374. report = raiw.remove_visible_detailed("watermarked.png", "clean.png")
  375. print(report.status) # cleaned | partial | unvalidated | no_watermark
  376. complete = raiw.remove_all("watermarked.png", "clean.png")
  377. print(complete.visible_status, complete.visible_marks)
  378. images = raiw.remove_batch("images", "images_clean", mode="visible")
  379. for item in images.items:
  380. print(item.source, item.visible_status, item.visible_marks)
  381. provenance = raiw.identify_video("input.mp4")
  382. report = raiw.inspect_video_metadata("input.mp4")
  383. complete = raiw.remove_video_all("input.mp4", "clean.mp4")
  384. print(complete.visual_invisible_action, complete.audio.watermark_status)
  385. batch = raiw.remove_video_batch("videos", "videos_clean")
  386. cleaned = raiw.remove_video_metadata("input.mp4")
  387. synthid_cleaned = raiw.remove_video_invisible("input.mp4", "synthid_clean.mp4")
  388. auto_cleaned = raiw.remove_video_invisible(
  389. "input.mp4", "auto_clean.mp4", profile="auto"
  390. ) # routes ByteDance/Seedance vs SynthID and prefers CUDA Wan 2.1 when available
  391. seedance_experiment = raiw.remove_video_seedance_experimental(
  392. "owned-label-cleaned.mp4", "seedance_clean.mp4", quality="delivery"
  393. ) # experimental pixel reconstruction; handle visible labels first
  394. seedance_complete = raiw.remove_video_all(
  395. "input.mp4", "seedance-clean.mp4",
  396. include_invisible=True, invisible_profile="seedance-experimental",
  397. ) # supported visible removal, then experimental reconstruction
  398. wan_synthid = raiw.remove_video_wan_experimental(
  399. "owned-veo.mp4", "veo_wan_clean.mp4", profile="wan21-synthid-experimental"
  400. ) # optional CUDA 3D Wan 2.1 video diffusion at native resolution (multi-window >193 frames)
  401. print(synthid_cleaned.visual_invisible_action, synthid_cleaned.audio.watermark_status)
  402. visible = raiw.remove_video_visible("input.mp4", "clean.mp4")
  403. print(visible.mark)
  404. veo = raiw.remove_video_visible("veo.mp4", "veo_clean.mp4", mark="veo")
  405. seedance = raiw.remove_video_visible(
  406. "seedance.mp4",
  407. "seedance_clean.mp4",
  408. mark="seedance",
  409. )
  410. dola = raiw.remove_video_visible("dola.mp4", "dola_clean.mp4", mark="dola")
  411. ```
  412. The high level API accepts a file path or a BGR NumPy array. For path inputs it
  413. also reads provenance metadata, preserves alpha, and can strip AI metadata from
  414. the written result.
  415. See the [Python API guide](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/python-api.md) for visible removal, the full
  416. `remove_all` and `remove_batch` pipeline, provenance inspection, metadata
  417. stripping, and diffusion usage.
  418. ## Agent skill
  419. Coding agents can install the published skill and drive this CLI:
  420. ```bash
  421. npx skills add wiltodelta/remove-ai-watermarks
  422. ```
  423. Claude Code can add this repository as a plugin marketplace:
  424. ```text
  425. /plugin marketplace add wiltodelta/remove-ai-watermarks
  426. /plugin install remove-ai-watermarks@remove-ai-watermarks
  427. ```
  428. Codex reads the same skill through [`.codex-plugin/plugin.json`](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/.codex-plugin/plugin.json).
  429. The skill refuses stock-agency and other third-party paid-asset marks. See
  430. [the agent skill guide](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/agent-skill.md).
  431. ## MCP server
  432. `remove-ai-watermarks mcp` serves `identify`, `remove` and `batch` as Model
  433. Context Protocol tools over stdio, each a thin wrapper around one CLI command:
  434. ```json
  435. {"mcpServers": {"remove-ai-watermarks": {"command": "uvx", "args": ["--python", "3.12", "--with", "remove-ai-watermarks[all,text-draft]", "remove-ai-watermarks", "mcp"]}}}
  436. ```
  437. See [the MCP server guide](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/mcp-server.md).
  438. <!-- mcp-name: io.github.wiltodelta/remove-ai-watermarks -->
  439. ## ComfyUI
  440. The separate
  441. [ComfyUI Remove AI Watermarks](https://github.com/wiltodelta/ComfyUI-remove-ai-watermarks)
  442. package provides nodes for visible removal, detection, region erasing, and
  443. invisible removal.
  444. ## Important limitations
  445. - A missing local signal means unknown, not clean. Proprietary pixel
  446. watermarks may remain after metadata has been stripped.
  447. [What each path leaves behind](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/known-limitations.md#what-each-path-leaves-behind)
  448. lists, per command, what may remain and how to check it.
  449. - Visible removal reconstructs a small region. Results depend on the background
  450. and selected fill backend.
  451. - Invisible removal changes the whole image and may alter faces, text, or fine
  452. detail.
  453. - Visible video removal recognizes the moving Sora 2 wordmark, the current Veo
  454. diamond plus legacy `Veo` text, the Seedance boxed `AI` label, and the fixed
  455. Doubao, Dola, Hailuo AI, Vidu, and Kling AI labels. It does not recognize the older Sora Turbo
  456. corner swirl or unregistered layouts from those providers.
  457. The classical OpenCV backend can smear structured backgrounds; use MI-GAN or
  458. LaMa when recovery quality matters.
  459. - Video SynthID regeneration changes resolution, frame rate, and image detail.
  460. The shipped profile is oracle-certified, but no public local decoder can
  461. certify an arbitrary output at runtime. Recheck unusually important outputs
  462. after provider changes.
  463. - Invisible-watermark removal through the CLI and high-level API requires CUDA.
  464. All diffusion profiles refuse any other device at construction rather than
  465. falling back to one that cannot run them. The specialized local Paint API,
  466. visible removal, metadata stripping and `identify` run without CUDA.
  467. - Provider watermark systems can change. Validate important outputs with the
  468. provider's own verifier when one is available.
  469. The shipped `video invisible` command uses the certified `noise_std=0.15`
  470. profile. The companion `scripts/video_synthid_sweep.py` research harness builds
  471. a matched re-encode control plus VAE-regenerated candidates and leaves the
  472. verifier verdict blank:
  473. ```bash
  474. uv run --extra video --extra diffusion python scripts/video_synthid_sweep.py input.mp4 -o sweep/
  475. ```
  476. To score what an output actually cost, use `scripts/video_fidelity_probe.py`:
  477. the engine's own PSNR is measured after the input resize and before final
  478. encoding, so only the probe sees the delivered picture.
  479. ```bash
  480. uv run --extra video python scripts/video_fidelity_probe.py input.mp4 input_clean.mp4
  481. ```
  482. The control must still be SynthID-positive before a negative candidate can
  483. count as removal evidence. In the 2026-07-29 two-clip calibration, both matched
  484. controls were positive in Gemini's built-in SynthID verifier; the stronger
  485. candidate was negative on both carriers, while a weaker candidate was
  486. negative on one. A later adversarial follow-up that asked ordinary Gemini to
  487. reinterpret the pixel result returned `UNAVAILABLE`; that follow-up was not a
  488. verifier rerun and does not invalidate the built-in verdicts. A 2026-07-31
  489. full-clip check on a public eight-second Veo sample found `0.10` still detected
  490. and `0.15` not detected, so `0.15` is now the certified default. The
  491. reproducible hashes and verdicts for that full-clip check -- one carrier, two
  492. rows -- live in `data/evaluations/video-synthid-oracle.csv`. The earlier
  493. two-clip calibration is narrative only: its verdicts were not recorded in a
  494. tracked manifest, so treat the certified default as resting on the 2026-07-31
  495. rows.
  496. ## Documentation
  497. Start with the [documentation index](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/index.md).
  498. - [Installation](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/installation.md)
  499. - [CLI guide](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/cli.md)
  500. - [Python API](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/python-api.md)
  501. - [Supported signals](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/supported-signals.md)
  502. - [Known limitations](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/known-limitations.md)
  503. - [Scope, safety, and legal notes](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/legal-and-safety.md)
  504. - [Module internals](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/module-internals.md)
  505. - [Release and distribution](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/release-and-distribution.md)
  506. - [Agent skill](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/agent-skill.md)
  507. Research notes and historical experiments are listed separately in the
  508. [documentation index](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/index.md). They explain past decisions but do not
  509. define the current public API.
  510. ## Contributing
  511. Install the development environment and run the project gate:
  512. ```bash
  513. uv sync --frozen --extra dev
  514. bash maintain.sh
  515. ```
  516. See [module internals](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/docs/module-internals.md) before changing a subsystem
  517. with documented invariants. Report vulnerabilities as described in
  518. [SECURITY.md](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/SECURITY.md).
  519. ## License
  520. [Apache 2.0](https://github.com/wiltodelta/remove-ai-watermarks/blob/main/LICENSE). Copyright 2025-2026 wiltodelta.
  521. Made by [Victor Kuznetsov](https://wiltodelta.com/).

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.