Complete Guide: How to Remove AI Watermarks from Gemini, DALL-E, Midjourney and More
Complete Guide: How to Remove AI Watermarks from Gemini, DALL-E, Midjourney and More
One Command to Strip All AI Watermarks
You just generated a perfect AI image, ready to use in your project or share on social media. But there's a problem — a Google Gemini sparkle logo in the corner, a "豆包AI生成" text strip across the bottom, or invisible C2PA metadata that tells every platform your image was AI-generated. In late May 2026, developer Victor K. open-sourced remove-ai-watermarks on GitHub — a unified CLI tool that handles visible watermarks, invisible watermarks, and AI metadata in one pass. It's already gathered 3,900+ stars in its first six weeks.
The Three Layers of AI Watermarks
Understanding what the tool does starts with understanding what's actually on your AI-generated images.
Visible Watermarks: These are the most obvious — the Google Gemini/Nano Banana sparkle logo (lower-right corner), ByteDance's Doubao "豆包AI生成" text strip (bottom-right), Jimeng's "★ 即梦AI" wordmark (bottom-right), Samsung Galaxy AI's locale-specific "✦ Contenuti generati dall'AI" strip (bottom-left). These are embedded via alpha blending, so cropping or blurring always damages the image.
Invisible Watermarks: You can't see them with the naked eye, but they're embedded at the pixel/spectral level — Google's SynthID (used in Gemini and Imagen), Stable Signature (used in Stable Diffusion 3 and SDXL), and TreeRing. SynthID operates in both the frequency and spatial domains, surviving screenshots and compression. Removal requires diffusion-based regeneration. There are now two versions: SynthID v1 (Gemini) and SynthID v2 (OpenAI's ChatGPT Images 2.0 / gpt-image-2).
Metadata Watermarks: Stored in file headers rather than pixels — C2PA Content Credentials, EXIF tags (prompt, seed, model hash, GPS data), XMP tags, IPTC "Made with AI" labels, PNG text chunks (ComfyUI and AUTOMATIC1111 parameter logs). These don't affect the image visually, but platforms like Instagram, X/Twitter, and Pinterest read them to auto-label content as "AI Generated."
Most watermark removers only handle the visible layer. remove-ai-watermarks is the first open-source tool to cover all three layers comprehensively.
What remove-ai-watermarks Can Do
Repository: https://github.com/wiltodelta/remove-ai-watermarks
License: Apache 2.0
Stars: 3.9k+
Latest: 46 releases
| AI Model | Visible | Invisible | Metadata |
|---|---|---|---|
| Google Gemini / Nano Banana / Gemini 3 Pro | ✅ Sparkle | ✅ SynthID v1+v2 | ✅ C2PA + EXIF |
| OpenAI DALL-E 3 / ChatGPT | – | – | ✅ C2PA |
| OpenAI ChatGPT Images 2.0 (gpt-image-2) | – | ✅ SynthID v2 + pixel WM | ✅ C2PA |
| Stable Diffusion / SDXL / FLUX | – | ✅ DWT-DCT (imwatermark) | ✅ PNG chunks / C2PA |
| Adobe Firefly | – | – | ✅ Content Credentials |
| Midjourney | – | – | ✅ EXIF + XMP |
| ByteDance Doubao / Jimeng | ✅ Text/Wordmark | – | ✅ TC260 + C2PA |
| Samsung Galaxy AI | ✅ Locale strip | – | ✅ C2PA |
| xAI Grok (Aurora) | – | – | ✅ EXIF signature |
| Meta AI | – | – | ✅ IPTC "Made with AI" |
Beyond image support, it handles video files (MP4/MOV/WebM) and audio files (MP3/WAV/FLAC/OGG) — stripping C2PA and metadata from container-level containers via ffmpeg.
Installation & Usage
Install
pip install remove-ai-watermarks
Requires Python 3.9+. For invisible watermark removal (SynthID), install GPU-dependent extras:
pip install remove-ai-watermarks[lama,controlnet,sdxl]
Step 1: Detect What Watermarks Are Present
remove-ai-watermarks identify input.jpg --json
The output shows watermark type, confidence score, origin model (Google, OpenAI, Stability AI, etc.), and C2PA issuer information — everything in one structured report.
Step 2: Remove Everything in One Command
remove-ai-watermarks all input.jpg -o clean.jpg
This single command:
- Detects and removes any visible watermark (sparkle, text strip, wordmark)
- Runs diffusion-based regeneration for invisible watermarks (SynthID, Stable Signature, TreeRing)
- Strips all metadata — C2PA manifests, EXIF, XMP, IPTC, PNG text chunks
- Outputs a clean image
Performance: Visible removal takes ~0.05s per image (no GPU needed). Invisible removal via diffusion takes ~50 denoising steps — 2-5 seconds on a GPU, slower on CPU.
Targeted Operations
Remove only visible watermarks:
remove-ai-watermarks remove-visible input.jpg -o clean.jpg
Strip metadata only:
remove-ai-watermarks strip-metadata input.jpg -o clean.jpg
Remove invisible watermarks only:
remove-ai-watermarks remove-invisible input.jpg -o clean.jpg
Python API
from remove_ai_watermarks import process
result = process("input.jpg", mode="all")
result.save("clean.jpg")
Batch Processing
Process an entire directory of images:
remove-ai-watermarks batch input/ output/ --mode all
How Visible Watermark Removal Works
The tool uses reverse-alpha blending. The developer captured clean alpha maps of each watermark under controlled conditions (black and gray backgrounds), then applied a three-stage normalized cross-correlation (NCC) detector to locate the watermark position and scale — even after the image was resized or cropped.
Formula: original = (watermarked − α × logo) / (1 − α)
For Doubao and Jimeng watermarks (which have unique textures per image), residual edge artifacts are cleaned via gradient-masked inpainting after the alpha reversal.
How Invisible Watermark Removal Works
The default pipeline uses SDXL + Canny ControlNet:
- VAE encode the image to latent space
- Forward diffusion (add noise)
- Denoise (~50 steps) guided by ControlNet to preserve text and face structure
- VAE decode back to pixel space
Vendor-adaptive strength — no manual flagging needed:
- OpenAI gpt-image → strength
0.20 - Google Gemini → strength
0.30 - Unknown source → strength
0.30
For inputs without text or faces, the pipeline automatically falls back to plain SDXL (no ControlNet), which is faster.
Small inputs (long side < 1024px) are auto-upscaled to 1024px via Lanczos or ESRGAN. Adaptive polish (on by default) restores detail after SynthID removal without re-introducing the watermark.
The developer evaluated every face-restoration approach available and deliberately chose not to include one — each method made the output "look more AI-generated." The raw ControlNet output is currently the least-AI-looking face state achievable.
Online Version (No CLI Needed)
If you prefer a web interface, visit raiw.cc (https://raiw.cc). The free tier handles visible watermark removal and metadata stripping. Invisible watermark removal (SynthID) uses cloud GPU and is billed per image.
Comparison with Traditional Methods
| Method | Suitable For | Quality | Difficulty | Speed |
|---|---|---|---|---|
| Crop / Blur | Corner visible watermarks | Damages image | Low | Instant |
| ExifTool metadata strip | C2PA / EXIF only | No pixel impact | Medium | Instant |
| JPEG re-compression | Weak invisible WMs | 10-20% removal | Medium | Seconds |
| SD img2img (ComfyUI) | Invisible WMs | Changes image | High | 1-3 min |
| remove-ai-watermarks | All types | Full 3-layer coverage | Low | 0.05s to seconds |
Use Cases
- Content creators: Prevent social platforms from auto-labeling AI-generated images in posts
- Digital artists: Clean tool traces from AI-assisted workflows before final presentation
- Developers and researchers: Batch-process large AI image datasets, removing redundant metadata
- Privacy-conscious users: Strip prompts, seeds, and hash values from image metadata before sharing
Ethical and Legal Considerations
The repository README explicitly states the tool is "intended for legal use only" and includes a detailed legal section. The Hacker News discussion reflects two valid perspectives:
Privacy perspective: Users should have control over their digital footprint — AI watermarks shouldn't function as irremovable tracking tags.
Misuse concern: Easy watermark removal makes it harder to distinguish AI-generated content from authentic photography, potentially worsening the misinformation landscape.
The key distinction: if you generated the image and want to remove markers for personal or portfolio use, that's generally accepted. Removing copyright watermarks from someone else's work and passing it off as your own violates copyright law and platform terms. Check your local regulations and platform policies before using these tools.
Summary
remove-ai-watermarks is currently the most comprehensive open-source AI watermark remover available — handling visible, invisible, and metadata watermarks across all major AI platforms in a single unified tool. With Apache 2.0 licensing, 3.9k+ GitHub stars, and both CLI and Python API options, it's accessible to everyone from casual users to batch-processing pipelines.
For anyone working regularly with AI-generated images, understanding these watermark technologies and removal capabilities also reveals where the AI content provenance ecosystem stands today — what markers can be stripped, what traces remain, and what this means for the future of AI content attribution.