gemini-watermark-remover

KO | EN

gitlite — search

#ai #typescript #golang #go #machine-learning #rust #docker #java #vue #pytorch #security #ios #mcp #openai

gemini-watermark-remover

(★ 1)

Accurately remove the visible Gemini/Nano-Banana sparkle watermark via reverse alpha-compositing. CLI + HTTP API.

gemini-watermark-remover Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Gemini Watermark Remover

Remove the **visible Gemini / Nano-Banana sparkle watermark** from images — *accurately*, by
inverting the exact alpha-compositing that applied it, with a template-free inpaint fallback so it
works out of the box. Ships as a **CLI** (`wmark`) and an **HTTP API**.

## Install

**From the GitHub release (recommended):**
```bash
python3 -m pip install "https://github.com/realvidhaan/gemini-watermark-remover/releases/latest/download/gemini_watermark_remover-0.1.0-py3-none-any.whl"
```
That installs the `wmark` command. If your shell can't find `wmark` afterward, either restart the
terminal or run it as `python3 -m wmark.cli remove <image>`. (Prefer isolated installs? Use
`pipx install <same url>` — but that needs pipx installed first: `brew install pipx`.)

**From source:**
```bash
git clone https://github.com/realvidhaan/gemini-watermark-remover
cd gemini-watermark-remover
pip install .            # or: pip install -e .   for development
```

Optional best-quality inpaint fallback: `pip install simple-lama-inpainting`.

## CLI usage

```bash
wmark remove path/to/image.png                 # writes image-clean.png next to it
wmark remove path/to/image.png -o out.png       # choose the output
wmark remove path/to/folder -o out/             # batch a whole folder
```

`remove` defaults to `--method auto`: it tries the **exact** reverse-alpha path and, if no trained
template removes the mark, falls back to **template-free inpainting with automatic sparkle
detection** — so a single image just works, no setup.

```bash
wmark remove img.png --method exact             # exact only (needs a trained template)
wmark remove img.png --method inpaint           # force inpaint
wmark remove img.png --box 386 466 36 42        # give the watermark box yourself
wmark remove img.png --sensitivity 0.6          # mask more (if some sparkle survives)
```

## Why the exact path is accurate

The sparkle is a **fixed, semi-transparent overlay** composited as:

```
watermarked = α · logo + (1 − α) · original
```

`logo` and `α` are the *same every time* for a given image size, so every image inverts exactly:

```
original = (watermarked − α · logo) / (1 − α)
```

This recovers the **original pixels** — no hallucination (**~48–54 dB PSNR** in tests). Generic
removers can't match it because they inpaint a blob instead of undoing the known compositing.

Only the **visible** sparkle is in scope. **SynthID** (the invisible, baked-in watermark) is *not*
touched — the image remains detectable as AI-generated.

## Unlock the exact path: train a template

The exact method needs a real `(logo, α)` template estimated from your own Gemini images (the CLI
works without it via inpaint). Collect 20–50 Gemini images per size; ideally with clean
counterparts (Ultra-tier / API outputs of the same prompts) for the exact **paired** solver:

```bash
wmark train watermarked_folder --clean clean_folder --bucket 1024   # exact
wmark train watermarked_folder --blind --bucket 1024                 # watermarked-only, approximate
```

Templates are stored per size bucket. Set `WM_TEMPLATES_DIR` to control where.

### Try it with no real images (synthetic demo)

```bash
python tools/make_synthetic.py   # demo templates + clean/watermarked pairs
python tools/evaluate.py         # PSNR/SSIM of recovery vs ground truth
pytest                           # near-lossless recovery + never damages clean images
```

## HTTP API

```bash
wmark serve --port 8000          # or:  uvicorn wmark.main:app --reload
```

- `POST /remove` — exact reverse-alpha. Multipart `image`; flags `force_inpaint`, `debug`.
- `POST /inpaint` — template-free. Multipart `image`; optional box `x,y,w,h` (auto-detects if omitted), `sensitivity`.
- `GET /health`

```bash
curl -F image=@photo.png "http://localhost:8000/remove"  -o out.png
curl -F image=@photo.png "http://localhost:8000/inpaint?x=386&y=466&w=36&h=42" -o out.png
curl -F image=@photo.png "http://localhost:8000/inpaint" -o out.png   # auto-detect
```

## Deploy (CPU is enough)

```bash
docker build -t gemini-wm-remover .
docker run -p 8000:8000 gemini-wm-remover
```

## Limits & honesty

- **"Exact" holds for un-recompressed inputs** from Gemini. JPEG re-saves make reverse-alpha
  approximate; `force_inpaint=true` cleans the residual with LaMa (26 → 42 → 47 dB in tests).
- **Auto-detection** uses a star-shaped matched filter; on cluttered corners it can miss — pass
  `--box` for certainty.
- **SynthID remains.** Removing Google's watermark violates Gemini's ToS and strips a provenance
  signal — fine for personal use, risky if passing AI images off as human-made.

## Layout

```
wmark/   template_store, localize, reverse_alpha, residual, inpaint, detect,
         corner_inpaint, remover, estimate, cli, main (FastAPI)
tools/   make_synthetic (demo), estimate_template (wrapper), evaluate (PSNR/SSIM)
tests/   test_reverse_alpha (near-lossless recovery + no-damage contract)
```

gemini-watermark-remover

gemini-watermark-remover Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

gemini-watermark-remover

gemini-watermark-remover Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README