xhs-anti-detection Skill

Purpose

Post-processing skill for Xiaohongshu (小红书) AI-generated images to reduce detection probability while maintaining visual quality. Applies a multi-layer defense strategy: metadata cleaning, subtle pixel modifications, and re-encoding.

When to Use

• After generating images with image-generation skill
• Before publishing to Xiaohongshu to avoid AI-generated content flags
• Batch processing multiple images at once
• When images need to appear "natural" or "camera-captured"

What It Does

Processing Pipeline

Input Image → Metadata Cleaner → Subtle Noise Adder → Color Shift → 
Text Area Protection → Re-compression → Verification → Safe Output

Layer 1: Metadata Cleaning

• Removes EXIF fields that reveal AI generation (Software, Creator, CreationDate)
• Fakes camera metadata (Canon EOS R5 / Sony A7M4 / iPhone 15 Pro)
• Preserves essential fields (dimensions, color profile)

Layer 2: Pixel-Level Modifications

• Adds Gaussian noise (σ=0.3, visually imperceptible)
• Applies subtle color shift (hue ±1°, saturation ±2%)
• Introduces micro-variations to break compression fingerprints

Layer 3: Text Protection

• Detects text regions (OCR-based)
• Applies sharpening to text areas (avoid blur)
• Leaves text crisp while background gets subtle processing

Layer 4: Re-encoding

• Re-compresses with libjpeg-turbo at 98% quality
• Shuffles DCT coefficient order
• Adds random padding bytes to break statistical patterns

Layer 5: Verification

• Checks metadata cleanliness
• Computes "naturalness" score
• Generates compliance report

Usage

Basic Usage

# Process single image
bash /Users/tianqu/.deskclaw/nanobot/workspace/skills/xhs-anti-detection/scripts/process.sh \
  --input /path/to/input.png \
  --output /path/to/output.png

# Batch process directory
bash /Users/tianqu/.deskclaw/nanobot/workspace/skills/xhs-anti-detection/scripts/batch.sh \
  --input-dir /path/to/images \
  --output-dir /path/to/processed

Parameters

| Flag | Description | Default | |------|-------------|---------| | --input | Input image path | (required) | | --output | Output image path | (required) | | --strength | Processing intensity: light/medium/heavy | medium | | --fake-camera | Camera model to fake | "Canon EOS R5" | | --verify | Run verification after processing | true | | --dry-run | Show what would be done without doing it | false |

Integration with image-generation

After generating an image with the image-generation skill, automatically run:

# Get the generated image path from image-generation output
# Then process it
bash ~/.deskclaw/nanobot/workspace/skills/xhs-anti-detection/scripts/process.sh \
  --input "$GENERATED_IMAGE" \
  --output "$GENERATED_IMAGE".safe.png \
  --strength medium

Configuration

Edit references/safe_params.json to adjust:

{
  "noise_sigma": 0.3,
  "color_shift_hue_deg": 1,
  "color_shift_saturation_pct": 2,
  "recompression_quality": 98,
  "text_sharpening_radius": 1,
  "metadata_fields_to_remove": [
    "Software", "Creator", "CreationDate", "DateTime",
    "Artist", "Copyright", "ExifVersion"
  ],
  "fake_camera_models": [
    "Canon EOS R5",
    "Sony A7M4",
    "iPhone 15 Pro",
    "Xiaomi 14 Ultra"
  ]
}

Output

• Processed image: Safe for publishing
• Verification report (if --verify): JSON with scores and warnings
• Original preserved: Input file is not modified

Limitations

• Not 100% guaranteed: Detection algorithms evolve continuously
• Slight quality loss: ~2-5% perceptible degradation (usually unnoticeable)
• Processing time: 3-5 seconds per image
• Text legibility: Text remains readable but may lose perfect crispness

Maintenance

• Update references/detection_patterns.md when new AI detection features are discovered
• Adjust safe_params.json if Xiaohongshu changes detection strategy
• Test with a burner account regularly

Files

xhs-anti-detection/
├── SKILL.md              # This file
├── scripts/
│   ├── process.sh        # Main entry point (bash wrapper)
│   ├── clean_metadata.py # EXIF cleaner
│   ├── add_noise.py      # Gaussian noise adder
│   ├── color_shift.py    # Subtle color modification
│   ├── protect_text.py   # Text-aware processing
│   ├── recompress.py     # Re-encoding with fingerprint randomization
│   ├── verify.py         # Verification & reporting
│   └── batch.sh          # Batch processing wrapper
├── references/
│   ├── safe_params.json  # Tunable parameters
│   └── detection_patterns.md  # Known detection signatures
├── hooks/
│   └── post_generate.py  # Auto-trigger after image-generation
└── assets/
    └── sample_report.json  # Example verification output

Dependencies

• Python 3.9+
• Pillow (PIL)
• pyexiv2 or exiftool (for metadata)
• numpy
• OpenCV (optional, for text detection)

Install:

pip install Pillow pyexiv2 numpy opencv-python

Future Enhancements

• [ ] Machine learning-based text region detection (more accurate)
• [ ] Adaptive strength based on image content complexity
• [ ] Automatic parameter tuning via A/B testing
• [ ] Integration as a post-processing hook for image-generation skill
• [ ] Support for video frames (extract → process → recompile)