DEV.co
AI Frameworks · wiltodelta

remove-ai-watermarks

Remove-AI-Watermarks is a Python CLI and library that strips visible and invisible AI watermarks (SynthID, Gemini sparkle, C2PA metadata) from AI-generated images. It handles multiple AI platforms and includes batch processing, with both local GPU and cloud-based processing options via raiw.cc.

Source: GitHub — github.com/wiltodelta/remove-ai-watermarks
3.9k
GitHub stars
352
Forks
Python
Primary language
Apache-2.0
License (OSI-approved)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Repositorywiltodelta/remove-ai-watermarks
Ownerwiltodelta
Primary languagePython
LicenseApache-2.0 — OSI-approved
Stars3.9k
Forks352
Open issues2
Latest releasev0.13.0 (2026-07-07)
Last updated2026-07-07
Sourcehttps://github.com/wiltodelta/remove-ai-watermarks

What remove-ai-watermarks is

Python-based tool using reverse-alpha blending for visible watermark removal, diffusion-based regeneration (SDXL/Qwen/ControlNet) for invisible watermarks, and metadata stripping (EXIF, XMP, C2PA, IPTC) across multiple image and media formats. Requires optional GPU for invisible watermark removal; metadata/visible removal runs on CPU.

Quickstart

Get the remove-ai-watermarks source

Clone the repository and explore it locally.

terminalbash
git clone https://github.com/wiltodelta/remove-ai-watermarks.gitcd remove-ai-watermarks# follow the project's README for install & configuration

Need it deployed, integrated, or customized instead? DEV.co ships production installs.

Best use cases

Removing platform-applied AI provenance marks from self-generated content

Strip AI-disclosure watermarks (SynthID, Gemini sparkle, C2PA labels) that platforms auto-apply to your own generated images, restoring full ownership and flexibility for distribution.

Batch cleaning AI-generated image archives

Process directories of AI images in one command to remove all detectable watermarks and metadata, useful for preparing bulk content for downstream applications or archiving.

Integrating AI watermark detection into image provenance workflows

Use the `identify` command and library API to detect and inventory AI-generated content, watermark type, and metadata source (C2PA issuer, AI system, visible marks) for compliance or content classification.

Implementation considerations

  • Visible watermark removal (sparkle, text marks) runs on CPU and is fast; invisible watermark removal requires CUDA GPU (~2 GB VRAM) or cloud GPU via raiw.cc (paid per image).
  • Multiple pipeline options (SDXL, Qwen, ControlNet) trade speed vs. fidelity; ControlNet is default and preserves faces/text better than plain SDXL, but Qwen is better for CJK text at cost of larger model.
  • Metadata stripping works offline and supports 10+ formats (PNG, JPEG, AVIF, HEIF, JPEG-XL, MP4, MOV, WebM, MP3, WAV, FLAC, OGG); requires ffmpeg for some formats.
  • Detection uses three-stage NCC scoring with confidence output; `identify` aggregates 15+ watermark/metadata signals into a single origin verdict (`--json` for parsing).
  • Face identity is not preserved by design—regenerated faces drift in likeness to avoid re-introducing SynthID; no shipped face-restore extra, by evaluation choice.

When to avoid it — and what to weigh

  • Removing stock agency, copyright, or purchase-gated watermarks — Tool explicitly excludes watermarks protecting paid/copyrighted content (Shutterstock, Getty, Adobe Stock). Use is out of scope by design; this is for AI-provenance marks, not anti-piracy.
  • Jurisdiction restricts AI-label removal — Some jurisdictions legally restrict removing AI-generated labels. Verify local law compliance before deployment; tool includes legal disclaimers but does not provide jurisdictional guidance.
  • Sensitive image processing without explicit user consent — Tool is intended for lawful use on content you generated yourself. Deploying on third-party images without consent, or for deception/fraud, violates stated scope and potentially local law.
  • Invisible watermark removal without GPU or cloud access — SynthID and other invisible watermark removal requires ~2 GB of local GPU VRAM or cloud GPU. If neither is available, metadata/visible-mark removal still works, but invisible watermark stripping is blocked.

License & commercial use

Licensed under Apache License 2.0 (Apache-2.0), a permissive OSI license. Allows commercial use, modification, and distribution with attribution and liability waiver. No restrictions on downstream use of the software itself, though users remain responsible for compliance with local law regarding removal of AI labels.

Apache-2.0 is permissive and permits commercial use, sale, and modification without royalties. However, the tool's scope disclaims removal of copyright/paid watermarks and notes that some jurisdictions restrict AI-label removal. Consult legal counsel on jurisdictional compliance before commercializing a service. The hosted raiw.cc charges per-image for invisible watermark removal (cloud GPU cost model), indicating commercial operation is feasible.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationStrong
License clarityClear
Deployment complexityModerate
DEV.co fitGood
Assessment confidenceHigh
Security considerations

Tool removes cryptographic watermarks and metadata; no claims of cryptographic security are made. Local operation avoids data transmission (unless cloud GPU used). Metadata stripping is non-reversible. C2PA manifests are removed (not verified); no integrity validation of stripped content. Model safety (SDXL, Qwen) is not audited; diffusion regeneration may produce unexpected outputs. Use on sensitive/medical images not discussed.

Alternatives to consider

Adobe Firefly / Content-Aware Fill

Commercial, closed-source alternative for region-based object removal; proprietary AI preservation. No watermark detection/stripping, and tied to Adobe ecosystem.

OpenCV inpainting + cv2.inpaint or Lama (standalone)

Generic region erasers without AI-watermark detection or metadata stripping. Requires manual specification of removal regions; no automated watermark detection or provenance inventory.

ffmpeg + exiftool

Lightweight CLI tools for metadata stripping only (EXIF, IPTC, XMP). No watermark detection or diffusion-based removal; purely metadata-focused.

Software development agency

Build on remove-ai-watermarks with DEV.co software developers

Review the code, test locally on sample images, and assess GPU/cloud GPU requirements. Confirm jurisdictional compliance before commercial deployment. Consider raiw.cc cloud option to avoid local GPU infrastructure.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

remove-ai-watermarks FAQ

Can I use this to remove watermarks from stock photos or copyrighted images?
No. The tool explicitly excludes watermarks protecting paid/copyrighted content (Shutterstock, Getty, Adobe Stock). It targets only AI-provenance marks (SynthID, sparkles, C2PA) on content you generated yourself. Removing stock watermarks is out of scope.
Do I need a GPU to remove invisible watermarks locally?
Yes, for SynthID/StableSignature/TreeRing removal via diffusion regeneration. ~2 GB CUDA GPU VRAM required. Alternatively, use raiw.cc cloud GPU for a per-image fee. Visible watermark and metadata removal work on CPU.
What happens if I remove a watermark and later my jurisdiction makes it illegal?
Tool includes legal disclaimers stating 'responsibility for any downstream use, and for compliance with local law, rests entirely with the user.' Consult legal counsel on your jurisdiction before use. Some regions restrict AI-label removal.
Does this tool re-introduce SynthID or other watermarks after removal?
No. Regenerated content is treated as new, not re-watermarked. The ControlNet/SDXL/Qwen pipeline explicitly avoids face-restore techniques that would re-introduce SynthID. Face identity drifts by design.

Custom software development services

From first prototype to production, DEV.co delivers software development services around tools like remove-ai-watermarks. Our software development agency staffs experienced software developers and web developers for custom software development, web development, integrations, and ongoing support across ai frameworks and beyond.

Evaluate Watermark Removal for Your Workflow

Review the code, test locally on sample images, and assess GPU/cloud GPU requirements. Confirm jurisdictional compliance before commercial deployment. Consider raiw.cc cloud option to avoid local GPU infrastructure.