Strip EXIF, GPS, C2PA / Content Credentials, AI watermarks, and Stable Diffusion / ComfyUI / Midjourney / ChatGPT prompt blocks from any image — fully offline.
A privacy-first desktop app + CLI that removes every identifying fingerprint embedded in your images, with first-class support for the AI provenance markers that ship inside images generated by ChatGPT, DALL·E, Midjourney, Adobe Firefly, Stable Diffusion, and ComfyUI.
Images you share online are loaded with identifiers most people never see:
- EXIF — your camera serial number, lens serial, capture date, and GPS coordinates of your home.
- C2PA / Content Credentials — cryptographically-signed provenance manifests that link the file back to OpenAI, Adobe, Google, or Leica. Increasingly required by law (EU AI Act, China generative-AI labelling rules) and shipped by default in every ChatGPT image since 2024.
- AI generation prompt blocks — Stable Diffusion A1111 writes your full prompt, negative prompt, seed, model hash, and sampler into a PNG
tEXtchunk. ComfyUI embeds the entire node graph as JSON. Midjourney puts your prompt inDescription. - IPTC
DigitalSourceType = trainedAlgorithmicMedia— the IPTC standard "this is AI-generated" tag. - Photoshop edit history, embedded thumbnails (often the unedited original), JPEG quantization-table fingerprints, JFIF/MPF/Adobe APP segments, and trailing payloads appended after end-of-file markers.
image-fingerprint-remover detects every category and offers three escalating cleaning modes — from a byte-identical metadata strip to an aggressive watermark-disruption pass.
- Comprehensive detector for PNG and JPEG containers, walking every chunk / segment at the byte level
- Three cleaning modes:
safe— strip metadata only, pixels byte-identical (verified via SHA256)paranoid— also re-encodes pixels with light Gaussian noise to neutralize JPEG quantization-table fingerprintsnuclear— also resize / crop / color-shift to disrupt frequency-domain robust watermarks (Google SynthID, Digimarc, IMATAG)
- Both CLI and native desktop GUI (PySide6, three-pane single-window app with drag-and-drop, parallel inspection, sequential cleaning with cancel support)
- Fully offline — no telemetry, no cloud uploads, no phone-home. Your images never leave your machine.
- JSON output for scripting and automation pipelines
- Honest about robust watermarks — we tell you what we cannot guarantee, rather than pretending
git clone https://github.com/lhfer/image-fingerprint-remover.git
cd image-fingerprint-remover
python3 -m venv .venv
.venv/bin/pip install Pillow numpy PySide6 # PySide6 only needed for the GUI# What identifiers does this image carry?
.venv/bin/python -m imgclean inspect photo.jpg
# Strip them (pixels unchanged)
.venv/bin/python -m imgclean clean photo.jpg -o cleaned.jpg
# Aggressive mode — also disrupts pixel-domain watermarks
.venv/bin/python -m imgclean clean photo.png -m nuclear -o cleaned.png
# Batch clean a folder
.venv/bin/python -m imgclean batch ./my-photos --recursive -m safe
# Compare two files (metadata diff + pixel SHA256)
.venv/bin/python -m imgclean diff original.png cleaned.png
# JSON output for any command
.venv/bin/python -m imgclean inspect image.png --jsonExit codes: 0 = clean / success · 1 = identifying metadata present (inspect) or residual after clean · 2 = usage / I/O error.
.venv/bin/python -m imgclean_guiDrag images or folders onto the window. Each file is inspected in parallel and rendered with its detection status and category badges. Pick a mode, click "清除 N 张图片", done.
| Category | Examples |
|---|---|
| C2PA / Content Credentials | OpenAI ChatGPT, Adobe Firefly, Google Imagen, Leica camera, Microsoft Designer JUMBF manifests in PNG caBX chunks or JPEG APP11 segments |
| AI generation prompts | Stable Diffusion A1111 parameters (prompt + negative + seed + model hash + sampler), ComfyUI prompt / workflow (full node graph), InvokeAI, NovelAI, Midjourney Description |
| IPTC AI tags | Iptc4xmpExt:DigitalSourceType = trainedAlgorithmicMedia / compositeWithTrainedAlgorithmicMedia |
| EXIF | Camera Make/Model, Software, DateTime, ImageDescription, UserComment, ImageUniqueID, HostComputer |
| GPS | EXIF GPS IFD — latitude, longitude, altitude |
| Device serials | BodySerialNumber, LensSerialNumber, SerialNumber |
| ICC color profiles | Embedded color profiles (can carry vendor IDs and custom strings) |
| Embedded thumbnails | EXIF IFD1 thumbnail, JPEG MPF, JFXX — often retain pre-edit pixels |
| Photoshop IRB | APP13 8BIM blocks — original filenames, paths, URLs, IPTC IIM |
| Adobe APP14 | DCT-transform marker (identifies Adobe encoder) |
| PNG text chunks | tEXt, iTXt, zTXt (including compressed) |
| PNG embedded EXIF | eXIf chunk |
| Unknown PNG chunks | Any non-standard ancillary chunk |
| JPEG quantization tables | Encoder/camera fingerprint that survives EXIF strip (LOW severity — only removed in paranoid/nuclear) |
| Trailing bytes | Bytes after IEND (PNG) or EOI (JPEG) — possible appended payload |
| Mode | What it does | Pixels bit-identical? | Best for |
|---|---|---|---|
| safe | Strip every non-essential metadata chunk/segment; drop trailing bytes; keep IDAT / scan data byte-for-byte. |
✅ | Posting images on the web while preserving exact quality. |
| paranoid | safe + re-encode pixels through a stock encoder + Gaussian noise σ=0.5 + reset filesystem mtime. | ❌ (imperceptible) | Neutralizing JPEG quantization-table fingerprints; bulk sanitization; weakens but does not defeat trained pixel watermarks. |
| nuclear | paranoid + resize to 99.7% + crop 2px + per-channel color bias + transcode through a different codec. | ❌ (visible-but-mild) | Best-effort disruption of robust frequency-domain watermarks (Google SynthID, Digimarc, IMATAG). |
imgclean/
imgclean/ # detection + cleaning engine
cli.py # argparse entrypoint
detect/
png.py # PNG chunk walker
jpeg.py # JPEG marker walker + EXIF/XMP/C2PA parsers
clean/
png.py # safe / paranoid / nuclear cleaners
jpeg.py # same, for JPEG
findings.py # Finding / InspectReport types + severity model
fingerprints.py # signature database (AI keys, C2PA labels, EXIF tags)
report.py # human-readable + JSON formatters
imgclean_gui/ # PySide6 desktop app
app.py # MainWindow + custom list delegate
workers.py # InspectTask (QThreadPool) + CleanWorker (QThread)
tests/ # 12 end-to-end tests covering each finding category
scripts/ # GUI smoke test, demo image generator, screenshot maker
We are honest about what we cannot guarantee:
- Robust pixel-domain watermarks (Google SynthID, Digimarc, IMATAG, Steg.AI) are co-trained against crop / resize / JPEG / color shift. The
nuclearmode disrupts them but does not guarantee removal. There is no public SynthID-Image detector to verify removal against. - C2PA durable credentials that include a server-side hash lookup survive any local sanitization — the originating service can still match by content hash even after the on-disk manifest is gone. Local stripping defeats the manifest-on-disk binding only.
- MakerNote in Canon / Nikon / Sony JPEGs is partially proprietary. We strip the whole block but do not enumerate every internal tag.
- HEIC / WebP / AVIF / TIFF / GIF containers are recognized but not yet fully cleaned in this build. PNG and JPEG are the supported targets.
- PRNU sensor noise is a pixel-statistics fingerprint and cannot be removed without destructive denoising.
If you need stronger guarantees against trained watermarks, the only known reliable approach is latent re-diffusion (encode → noise → denoise → decode through a local Stable Diffusion model). That is out of scope for this MVP.
.venv/bin/python -m pytest tests/ -v # engine tests
QT_QPA_PLATFORM=offscreen .venv/bin/python scripts/gui_smoke.py # headless GUI smokeThe engine ships with 12 end-to-end tests covering every finding category and asserting safe mode preserves pixels bit-identically.
图像指纹清理工具 — 检测并清除图像中所有可能携带的识别元素：
- EXIF / GPS / 设备序列号 / 拍摄时间
- C2PA 内容凭证清单（ChatGPT、Adobe Firefly、Google Imagen 等都会嵌入这种数字签名清单）
- AI 生成提示词（Stable Diffusion A1111 的
parameters、ComfyUI 的prompt/workflow节点图、Midjourney 的提示词块、Adobe Firefly 标签） - IPTC 标准的 AI 生成标签
- Photoshop 编辑历史 / 嵌入缩略图 / JPEG 量化表指纹 / 尾部追加数据
三档清理模式：安全（像素不变）、深度（重编码 + 轻噪声）、核弹（缩放裁剪 + 色彩微调，扰乱鲁棒水印）。提供命令行 + macOS 原生桌面应用（PySide6，支持拖拽、批量、并行检测），100% 本地运行，不联网。
Issues and PRs welcome. Particularly interested in:
- HEIC / WebP / AVIF / TIFF cleaning paths
- Better robust-watermark detection heuristics
- Linux / Windows GUI polish
- Localization beyond Chinese / English
MIT — see LICENSE.
- The C2PA / Content Authenticity Initiative for publishing the JUMBF spec openly
- Phil Harvey's ExifTool — the reference implementation for image metadata that every project in this space owes a debt to
- The Stable Diffusion and ComfyUI communities for documenting their PNG metadata conventions