reverse-synthid

name	reverse-synthid
description	Run Google's SynthID watermark detection and removal via the reverse-SynthID project on the user's main PC (ssh main). Use when the user mentions SynthID, watermark bypass, watermark removal, AI image watermark, Gemini watermark, or wants to run reverse-SynthID on images.

reverse-SynthID Skill

The user's main PC (accessible via ssh main) has the reverse-SynthID project cloned at F:\specprojects\reverse-SynthID. All dependencies (including PyTorch with CUDA) are already installed.

Quick Reference

Strength	Uses GPU	PSNR	Speed	Best For
gentle	No	~50 dB	~5 sec	Minimal disruption
moderate	No	~48 dB	~5 sec	Balanced spectral
aggressive	No	~45 dB	~5 sec	Strong spectral
maximum	No	~50 dB	~10 sec	Strongest non-VAE
final	Yes	~16 dB	~90 sec	Detector-defeating (7-stage)
nuke	Yes	~12 dB	~180 sec	Maximum attack (2× VAE)

Project Layout on Main PC

F:\specprojects\reverse-SynthID\
├── artifacts\
│   ├── spectral_codebook_v4.npz      # V4 codebook (use this)
│   └── spectral_codebook_v3.npz      # Legacy V3
├── src\extraction\
│   ├── synthid_bypass_v4.py          # V4 bypass engine
│   └── vae_regen.py                  # VAE roundtrip (GPU)
├── scripts\
│   └── dissolve_batch.py             # Batch processing CLI
└── test_output/                      # Default output dir

Batch Processing

Use dissolve_batch.py for all operations. It handles profile auto-selection, manifest generation, and variant naming.

Basic Command Pattern

ssh main "cd F:/specprojects/reverse-SynthID && python scripts/dissolve_batch.py \
    --input <PATH_OR_DIR> \
    --output <OUT_DIR> \
    --codebook artifacts/spectral_codebook_v4.npz \
    --strengths <STRENGTH> \
    --model nano-banana-pro-preview"

Models

• gemini-3.1-flash-image-preview — Gemini Flash images
• nano-banana-pro-preview — Nano Banana Pro images (default for this user)
• Omit --model for auto-selection by resolution

Strength Presets

Available: gentle, moderate, aggressive, maximum, demolish, annihilate, combo, blog_pure, blog_plus, blog_combo, residual_pure, residual_plus, residual_combo, regen_pure, regen_plus, regen_combo, final, nuke

Multiple strengths can be passed (e.g., --strengths gentle moderate aggressive) to generate variants A, B, C.

Example Workflows

Single image, maximum spectral (CPU-only, fast):

ssh main "cd F:/specprojects/reverse-SynthID && python scripts/dissolve_batch.py \
    --input D:/lora/gemini\\ gens/images/1765719471693-to4wyv.webp \
    --output test_output \
    --codebook artifacts/spectral_codebook_v4.npz \
    --strengths maximum --model nano-banana-pro-preview"

Full directory, final strength (GPU, detector-defeating):

ssh main "cd F:/specprojects/reverse-SynthID && python scripts/dissolve_batch.py \
    --input D:/lora/gemini\\ gens/images \
    --output test_output_final \
    --codebook artifacts/spectral_codebook_v4.npz \
    --strengths final --model nano-banana-pro-preview"

Multiple strength variants for manual testing:

ssh main "cd F:/specprojects/reverse-SynthID && python scripts/dissolve_batch.py \
    --input <PATH> --output test_variants \
    --codebook artifacts/spectral_codebook_v4.npz \
    --strengths gentle moderate aggressive final \
    --model nano-banana-pro-preview"

Output Files

For each input image, outputs are named: <basename>__<VARIANT>_<strength>.png

A manifest.csv is generated with per-image metrics (PSNR, SSIM, profile, elapsed time).

File Transfer Between Main PC and Local

Copy from main PC to local temp:

# For single file
ssh main "cd <DIR> && cmd /c type <FILENAME>" > /tmp/<localname>

# For multiple files
ssh main "cd F:/specprojects/reverse-SynthID/test_output && cmd /c type file1.png" > /tmp/file1.png

Copy local file to main PC:

# Create temp file locally, then pipe via SSH
# (Use base64 encoding for binary files if needed)

Creating Comparisons

To create side-by-side original vs bypassed comparisons, use Python with Pillow on the main PC:

from PIL import Image, ImageDraw, ImageFont

# Load original and bypassed
img1 = Image.open(orig_path).convert('RGB')
img2 = Image.open(bypassed_path).convert('RGB')

# Resize to same height
target_h = 1080
w1 = int(img1.width * target_h / img1.height)
w2 = int(img2.width * target_h / img2.height)
img1 = img1.resize((w1, target_h), Image.LANCZOS)
img2 = img2.resize((w2, target_h), Image.LANCZOS)

# Side-by-side
result = Image.new('RGB', (w1 + w2, target_h))
result.paste(img1, (0, 0))
result.paste(img2, (w1, 0))

# Add labels
font = ImageFont.truetype('C:/Windows/Fonts/arial.ttf', 36)
draw = ImageDraw.Draw(result)
draw.text((20, 20), 'ORIGINAL', fill='white', font=font, stroke_width=3, stroke_fill='black')
draw.text((w1 + 20, 20), 'BYPASSED', fill='white', font=font, stroke_width=3, stroke_fill='black')

result.save(out_path, quality=95)

GPU Status Check

If VAE stages are slow, verify CUDA:

ssh main "python -c 'import torch; print(torch.cuda.is_available()); print(torch.cuda.get_device_name(0))'"

If torch.cuda.is_available() is False, reinstall PyTorch with CUDA:

ssh main "python -m pip uninstall torch torchvision -y && python -m pip install torch torchvision --index-url https://download.pytorch.org/whl/cu126"

Common Issues

Issue	Fix
UnicodeEncodeError on →	Replace \u2192 with -> in dissolve_batch.py line ~169
Spaces in Windows paths	Quote with single quotes inside double quotes: "'D:/path/with spaces/file.webp'"
VAE download timeout	First run only — cached to ~/.cache/huggingface/hub/
Slow on CPU	Must use CUDA PyTorch; see GPU Status Check above

Calibration Loop (Advanced)

After running variants, test in Gemini app and record results:

1. Upload each output to Gemini, run SynthID detection
2. Create tally.csv with columns: source,variant,still_watermarked (y/n)
3. Run calibration:

ssh main "cd F:/specprojects/reverse-SynthID && python scripts/calibrate_from_feedback.py \
    --manifest <manifest.csv> --tally <tally.csv> \
    --codebook artifacts/spectral_codebook_v4.npz"