Watermark Remover

Remove watermarks from images using Florence-2 detection + IOPaint (LaMa) inpainting. Supports batch processing and manual/automatic modes.

Automatically detect and remove watermarks (especially MLS watermarks) from listing photos using Florence-2 for detection and IOPaint (LaMa) for inpainting.

Prerequisites

pip install iopaint transformers torch pillow
# Optional OCR fallback:
pip install paddleocr paddlepaddle

Models auto-download on first run (~560MB total): LaMa (~100MB) + Florence-2 (~460MB).

Quick Start

# Single image
python ~/.openclaw/workspace/skills/watermark-remover/scripts/remove_watermark.py \
  --input photo.jpg --output photo_clean.jpg

# Batch directory
python ~/.openclaw/workspace/skills/watermark-remover/scripts/remove_watermark.py \
  --input ./photos/ --output ./photos_clean/ --suffix _clean

Script: remove_watermark.py

• --input — file or directory
• --output — file or directory (created if missing)
• --suffix — append to output filenames (e.g. _clean)
• --model — lama (default), mat, migan, or ldm
• --device — cpu, cuda, or mps (auto-detected)
• --confidence — detection threshold 0.0–1.0 (default: 0.5)
• --padding — mask expansion in pixels (default: 10)
• --dry-run — detect only, skip inpainting
• --preserve-exif — copy EXIF metadata (default: on)

Supported: .jpg, .jpeg, .png, .webp, .tiff

Pipeline

1. Detect — Florence-2 object detection (or PaddleOCR fallback) finds watermark regions
2. Mask — Generate binary mask with padding around detected boxes
3. Inpaint — IOPaint LaMa fills masked region with contextually appropriate pixels

Model Selection

• LaMa (default) — Fast, excellent for small text watermarks. Handles ~90% of MLS watermarks.
• MAT — Same speed/quality, different artifacts. Try if LaMa isn't clean enough.
• MIGAN — Lightest (~30MB). For CPU-only or low-VRAM environments.
• LDM — Slow but highest quality. For complex textures (patterned carpet, wallpaper).

Start with lama. Switch to ldm only if LaMa leaves visible artifacts.

Troubleshooting

• Ghost text remains — increase --padding or try ldm
• Blurry patch — switch to ldm for complex backgrounds
• Watermark not detected — lower --confidence to 0.3
• OOM — use --device cpu or migan model
• Color mismatch — add --match-histograms (IOPaint ≥1.5)

Notes

• Original images never modified in-place
• Fully deterministic (LaMa is non-stochastic)
• MLS watermarks (VMLS, CRMLS, etc.) are ideal LaMa use case: small, corner-positioned, semi-transparent text

See references/model-comparison.md for detailed model benchmarks.