1. Workflow Overview

This workflow is designed for automated or semi-automated removal of watermarks and logos from images. It supports two different detection modules (Florence2 and GroundingDINO + SAM) and allows manual mask adjustments. The final output is a cleaned image with a comparison view.

2. Core Models

3. Key Nodes

(1) Image Loading & Preprocessing

• LoadImage: Load input image (e.g., watermark-image-1.png).

• ImageScaleByAspectRatio: Resize image (e.g., 1024x1024) while maintaining aspect ratio.

(2) Watermark Detection Modules

Module 1: Florence2 (OCR+Region Detection)

• Florence2ModelLoader: Load Florence2 model (requires HuggingFace token).

• Florence2Run: Run detection to output masks and text data.

Module 2: GroundingDINO + SAM

• GroundingDinoSAMSegment: Combines DINO (detection) and SAM (segmentation) for mask generation.

• GrowMaskWithBlur: Expands mask edges for natural blending.

(3) Image Inpainting

• InpaintCrop: Crop the region to be repaired.

• KSampler: Use Stable Diffusion (e.g., realisticVisionV60B1) for inpainting.

• InpaintStitch: Stitch the inpainted area back to the original image.

(4) Output & Comparison

• SaveImage: Save the final result.

• Image Comparer: Side-by-side comparison of before/after results.

4. Workflow Structure

Group 1: Input & Preprocessing

• Input: Image file (e.g., PNG/JPG).

• Output: Resized image.

Group 2: Watermark Detection (Dual Modules)

• Input: Resized image.

• Output: Mask (for watermark/logo).

Group 3: Inpainting & Stitching

• Input: Mask + original image.

• Output: Repaired image.

Group 4: Result Comparison

• Input: Original vs. repaired image.

• Output: Comparison slider view.

5. Inputs & Outputs

Input Parameters

1. Image file (e.g., watermark-image-1.png).

2. Model choice (Florence2 or GroundingDINO+SAM).

3. Inpainting settings (e.g., steps 20, CFG 1).

Output

• Cleaned image (watermark removed).

• Comparison view (optional).

6. Notes

Installation Dependencies

1. Florence2: Requires HuggingFace account and API token.

   pip install transformers

2. GroundingDINO + SAM:

   • Install from GitHub:

     git clone https://github.com/IDEA-Research/GroundingDINO
     git clone https://github.com/facebookresearch/segment-anything

   • Download model weights (e.g., sam_vit_h_4b8939.pth).

3. ComfyUI Plugins:

   • Install via ComfyUI Manager:

     • rgthree Nodes (for context management).

     • LayerUtility (image scaling).

Troubleshooting

• CUDA Out of Memory: Reduce resolution or launch ComfyUI with --medvram.

• Model Load Failures: Verify weight file paths.

• Inaccurate Masks: Manually adjust GrowMaskWithBlur (e.g., expand pixels=5).

Optimization Tips

• Use RTX 3090/4090 or better GPU for optimal speed.

• For 4K images, downsample to 1024x1024 first.

7. Summary

This workflow is ideal for batch watermark removal, combining auto-detection with manual control. Florence2 works best for text, while GroundingDINO+SAM handles graphic logos. Output quality depends on the SD model (e.g., realisticVision for photos, revAnimated for anime).

Tip: Test on small images first and fine-tune mask settings!