Skip to content

SamurAIGPT/seedance-2.0-watermark-remover

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

13 Commits
 
 
 
 
 
 

Repository files navigation

Seedance 2.0 Watermark Remover

Remove the Seedance 2.0 watermark ("AI生成" / AI-Generated) from videos automatically — no GPU required, no paid tools.

Works on all videos generated by Seedance 2.0, Seedance Pro, and Seedance Lite video generation models. Also removes watermarks from any corner-positioned logo or text overlay.

Free • Open Source • No GPU Needed • Works on CPU

Don't want to run code locally? Use the free cloud version at muapi.ai/playground/seedance-2.0-watermark-remover — upload your video and get a clean version instantly.


What Is the Seedance 2.0 Watermark?

Seedance 2.0 (by ByteDance) adds a small "AI生成" (AI-Generated) badge to the corner of every generated video. This tool detects and removes that watermark automatically, restoring a clean video without artifacts.


How It Works

  1. Sample & average ~60 frames — the static Seedance watermark becomes clearly visible in the mean frame while moving content blurs away
  2. Auto-detect corner using Canny edge density × temporal stability score — the watermark is static and structured; moving content (people, water, foliage) scores low and is ignored
  3. Build a precise mask via Canny edge detection on the watermark region — traces only the text strokes
  4. Inpaint each frame with OpenCV TELEA (default, CPU-only) or optionally LaMa AI inpainting (--lama)
  5. Reassemble all frames with the original audio using ffmpeg

Installation

# Core dependencies (required)
pip install opencv-python-headless numpy

# ffmpeg (required for video reassembly)
# macOS
brew install ffmpeg
# Ubuntu / Debian
sudo apt install ffmpeg

# Optional: LaMa AI inpainting (--lama flag)
pip install torch iopaint

Usage

# Remove Seedance 2.0 watermark — auto-detects corner
python watermark_remover.py input.mp4

# Save to custom output path
python watermark_remover.py input.mp4 -o clean.mp4

# Manual region if auto-detection fails (x, y, width, height in pixels)
python watermark_remover.py input.mp4 -r 10,5,120,60

# Use LaMa AI inpainting for higher quality output (requires torch + iopaint)
python watermark_remover.py input.mp4 --lama

Features

  • Removes Seedance 2.0 watermark ("AI生成") automatically
  • Works on portrait and landscape video orientations
  • Detects watermark in any of the four corners
  • Handles videos where people or content move near the watermark corner
  • Supports both opaque and semi-transparent watermarks
  • No GPU required — runs entirely on CPU with OpenCV TELEA inpainting
  • Preserves original audio in output
  • Optional LaMa AI inpainting for higher quality results

How Auto-Detection Works

Each corner region (8% height × 12% width) is scored:

score = edge_density × (1 / (1 + temporal_std))
Term Meaning
edge_density Fraction of Canny edge pixels in the mean frame — watermark text has crisp, consistent edges
temporal_std Pixel variation across frames — moving content scores high (bad), static watermark scores low (good)

Using tight corner regions prevents a moving person near the corner from masking the watermark signal.


Options

Flag Description
input Path to input video
-o, --output Output file path (default: <input>_clean.mp4)
-r, --region Manual watermark region x,y,w,h — skips auto-detection
--lama Use LaMa AI inpainting (requires torch + iopaint)

Requirements

  • Python 3.8+
  • opencv-python or opencv-python-headless
  • numpy
  • ffmpeg (system install)
  • torch + iopaint (only for --lama)

Related

  • Free Cloud Version — remove Seedance watermarks online, no setup required
  • muapi.ai — API platform for Seedance 2.0, Seedance Pro, and other generative media models
  • Seedance 2.0 — ByteDance video generation model
  • iopaint — LaMa inpainting library

License

MIT