[English] | 日本語

---

ComfyUI Save Image with Watermark 💧

Custom node for ComfyUI with watermarking capabilities

🎨 This Image Has a Secret

graph TB
    subgraph visible["👁️ Visible Layer"]
        direction LR
        V1["🖼️ Main Image"]
        V2["🅰️ Logo Watermark<br/>(bottom-left)"]
        V3["📝 Text Watermark<br/>(bottom-right)"]
    end

    subgraph invisible["🔒 Invisible Layer - LSB Steganography"]
        direction LR
        I1["Pixel RGB Values"]
        I2["Embedded in<br/>Least Significant Bit"]
        I3["🔐 Hidden Message"]
        I1 --> I2 --> I3
    end

    IMG["📷 PNG Image File"]
    IMG --> visible
    IMG --> invisible

    style visible fill:#e8f5e9,stroke:#4caf50,stroke-width:2px
    style invisible fill:#fff8e1,stroke:#ff8f00,stroke-width:2px
    style I3 fill:#ffebee,stroke:#e53935

💡 Try it yourself!

cd examples
python3 -m venv venv && source venv/bin/activate && pip install Pillow
python lsb_demo.py aicuty_000011.png
# → Extracts "invisible Hello World!"

Features

Image Logo Watermark

• MASK Support: Accurate alpha blending using LoadImage's MASK output
• Transparent PNG Support: No black/white bleeding in compositing
• Position, Scale, Opacity: Fine-grained control

Text Watermark

• Custom Text: Configurable font size, color, opacity, position
• Dynamic Text: Embed seed values etc. from external nodes
• Tile Mode: Repeat pattern across entire image

Invisible Watermark (Steganography)

• LSB Embedding: Hide messages in the least significant bits
• Extraction Node: Extract embedded hidden messages

Metadata & Provenance

• Workflow Embedding: ComfyUI Core compatible workflow saving
• Content Hash: Generate SHA-256 hash for blockchain provenance
• AICU Metadata: Watermark info, timestamps, etc.

Save Options

• output_folder: Save to ComfyUI output folder
• browser_download: Download via browser
• both: Both (default)

Installation

Via ComfyUI Manager (Recommended)

1. Launch ComfyUI
2. Click Manager button
3. Select Install Custom Nodes
4. Search for watermark or aicu
5. Find comfyui-save-image-watermark and click Install
6. Restart ComfyUI

Via Comfy Registry

comfy node registry-install comfyui-save-image-watermark

Manual Installation

cd ComfyUI/custom_nodes
git clone https://github.com/aicuai/comfyui-save-image-watermark.git

Dependencies

This node uses only ComfyUI's standard libraries, no additional dependencies required.

• Pillow (bundled with ComfyUI)
• NumPy (bundled with ComfyUI)
• PyTorch (bundled with ComfyUI)

Node List

Save Image (Watermark) 💧

Category: AICU/Save

Main node for saving images with watermarks.

Extract Hidden Watermark 🔍

Category: AICU/Watermark

Node for extracting invisible watermarks (steganography).

---

How to Use in ComfyUI

Adding the Node

1. Right-click on canvas → Add Node
2. Select AICU → Save → Save Image (Watermark) 💧

Node Wiring Diagram

flowchart LR
    subgraph input["Input"]
        KS[KSampler] --> VAE[VAEDecode]
        LI[LoadImage<br/>Logo Image]
    end

    subgraph main["Save Image (Watermark) 💧"]
        direction TB
        I1[/"images (IMAGE)"/]
        I2[/"watermark_image (IMAGE)"/]
        I3[/"watermark_image_mask (MASK)"/]
        I4[/"dynamic_text (STRING)"/]

        O1[\"image (IMAGE)"\]
        O2[\"filename (STRING)"\]
        O3[\"content_hash (STRING)"\]
    end

    subgraph output["Output"]
        PREVIEW[PreviewImage]
        NEXT[Next Node...]
    end

    VAE -->|IMAGE| I1
    LI -->|IMAGE| I2
    LI -->|MASK| I3

    O1 --> PREVIEW
    O1 --> NEXT

Basic Connections

Step 1: Connect Source Image

Connect your generated image to the images pin.

[KSampler] → [VAEDecode] → images

Step 2: Connect Logo Image (Optional)

To add a logo watermark, connect two wires from LoadImage node.

[LoadImage]
    ├─ IMAGE → watermark_image
    └─ MASK  → watermark_image_mask  ← Important! Alpha info

⚠️ Don't forget MASK! Without MASK, transparent areas will appear black.

Step 3: Configure Parameters

Adjust settings using the node's widgets.

Input Pins (Left Side)

Pin Name	Type	Required	Description
images	IMAGE	✅	Source image. Connect VAEDecode output
watermark_image	IMAGE	-	Logo image. Connect LoadImage IMAGE output
watermark_image_mask	MASK	-	Logo alpha. Connect LoadImage MASK output
dynamic_text	STRING	-	Dynamic text. Can connect seed values etc.

Output Pins (Right Side)

Pin Name	Type	Description
image	IMAGE	Processed image. Can connect to subsequent nodes
filename	STRING	Saved filename
content_hash	STRING	SHA-256 hash (for provenance)

---

Parameter Details

Basic Settings

Parameter	Type	Default	Description
filename_prefix	STRING	"aicuty"	Filename prefix. Numbered like aicuty_00001_.png
file_format	ENUM	PNG	Save format. PNG recommended (preserves LSB)
save_to	ENUM	both	Destination. output_folder / browser_download / both

Image Logo Watermark

Parameter	Type	Default	Description
watermark_image_position	ENUM	bottom_left	Logo position
watermark_image_scale	FLOAT	0.15	Logo ratio to image width (1%-100%)
watermark_image_opacity	FLOAT	1.0	Opacity (0.0=transparent, 1.0=opaque)

Text Watermark

Parameter	Type	Default	Description
watermark_text	STRING	"© AICU"	Watermark text
watermark_text_enabled	BOOL	True	Enable/disable text watermark
watermark_text_position	ENUM	bottom_right	Text position
watermark_text_opacity	FLOAT	0.9	Opacity
watermark_text_size	INT	24	Font size (8-128px)
watermark_text_color	STRING	#FFFFFF	Text color (HEX format)

Invisible Watermark (LSB)

Parameter	Type	Default	Description
invisible_watermark	STRING	""	Secret message to embed
invisible_watermark_enabled	BOOL	False	Enable/disable invisible watermark

⚠️ Note: Invisible watermarks are preserved only in PNG format. JPEG/WebP will destroy them.

Metadata

Parameter	Type	Default	Description
embed_workflow	BOOL	True	Embed ComfyUI workflow
embed_metadata	BOOL	True	Embed AICU metadata
metadata_json	STRING	"{}"	Custom JSON metadata

Quality Settings

Parameter	Type	Default	Description
jpeg_quality	INT	95	JPEG quality (1-100)
webp_quality	INT	90	WebP quality (1-100)

---

Position Options

┌────────────────────────────────────────┐
│ top_left          top_right            │
│    ◆                    ◆              │
│                                        │
│              center                    │
│                 ◆                      │
│                                        │
│    ◆                    ◆              │
│ bottom_left      bottom_right          │
└────────────────────────────────────────┘

tile: Repeat pattern across entire image

---

Extract Hidden Watermark 🔍 Node

Node for extracting LSB steganography messages.

Input Pins

Pin Name	Type	Description
image	IMAGE	Image to extract from

Parameters

Parameter	Type	Default	Description
max_length	INT	1000	Maximum characters to read

Output Pins

Pin Name	Type	Description
hidden_message	STRING	Extracted message

Usage Example

[LoadImage] ──IMAGE──→ [Extract Hidden Watermark 🔍] ──STRING──→ [ShowText]

---

Processing Order

1. Image Logo Watermark (bottom layer)
   └─ Blend only in MASK areas with opacity

2. Text Watermark (above logo)
   └─ Combined with dynamic text

3. Invisible Watermark (last)
   └─ LSB Steganography

4. File Save

---

Technical Specifications

Image Logo Blending

• MASK=0: Fully transparent (no blending)
• MASK=255: Blend with opacity value
• Formula: result = base * (1 - mask * opacity) + logo * (mask * opacity)

Invisible Watermark (LSB)

This implementation uses simple LSB (Least Significant Bit) method.

What is LSB?

graph LR
    subgraph pixel["1 Pixel Structure"]
        R["R: 10000000<br/>(128)"]
        G["G: 10000000<br/>(128)"]
        B["B: 10000000<br/>(128)"]
    end

    subgraph lsb["Least Significant Bit (LSB)"]
        R --> R_LSB["0"]
        G --> G_LSB["0"]
        B --> B_LSB["0"]
    end

    R_LSB --> MSG["Message<br/>bit 1"]
    G_LSB --> MSG2["Message<br/>bit 2"]
    B_LSB --> MSG3["Message<br/>bit 3"]

    style R fill:#ffcdd2
    style G fill:#c8e6c9
    style B fill:#bbdefb

Example: Embedding "H" (ASCII 72 = 01001000)

Original Pixel      After Embedding     Change
─────────────────────────────────────────────
Pixel[0]
  R: 128 (10000000) → 128 (1000000[0])   LSB=0 ✓
  G: 128 (10000000) → 129 (1000000[1])   LSB=1 ← Changed!
  B: 128 (10000000) → 128 (1000000[0])   LSB=0 ✓

The 128 → 129 change is invisible to human eyes!

💡 Try it: python examples/lsb_demo.py --create-gray

Algorithm

Embedding Process:

1. Convert message to UTF-8 byte sequence
2. Add termination marker (4 NULL bytes \x00\x00\x00\x00)
3. Split each byte into 8 bits
4. Overwrite LSB of each pixel's R, G, B values sequentially
5. Alpha channel (transparency) is not modified

Extraction Process:

1. Get LSB from each pixel's R, G, B values sequentially
2. Collect 8 bits to reconstruct bytes
3. Stop when termination marker (4 consecutive NULL bytes) is detected
4. Decode byte sequence as UTF-8 to recover message

Specifications

• Maximum capacity: (width × height × 3) / 8 bytes
• Termination marker: 4 bytes (\x00\x00\x00\x00)
• Encoding: UTF-8

⚠️ Important Limitations

This method is vulnerable to image processing. The following operations will destroy data:

Operation	Effect
JPEG save	❌ Completely destroyed (lossy compression changes LSB)
WebP save (lossy)	❌ Completely destroyed
Resize	❌ Completely destroyed (pixel interpolation changes LSB)
Crop	❌ Completely destroyed (pixel positions shift)
Rotation	❌ Completely destroyed
Color adjustment	❌ Completely destroyed
PNG re-save	✅ Preserved (lossless compression)

Recommendations:

• Use PNG format only when using invisible watermarks
• Do not process the image before extraction
• For robustness, wait for future DCT/DWT implementation

Security Notice

• Current implementation has no encryption
• Embedding position is fixed pattern (sequential from top-left)
• Third parties can extract using the same algorithm
• Not suitable for embedding confidential information

Content Hash

• Algorithm: SHA-256
• Input: Image byte sequence encoded as PNG
• Output: 64-character hexadecimal string

---

Experiment: Does LSB Survive Photoshop Editing?

LSB steganography is said to be "vulnerable to image processing", but what actually happens when editing with Photoshop?

Test Environment

• Adobe Photoshop 2026
• macOS
• Source: 128x128 gray image (with "Hello LSB!" embedded via LSB)

Procedure

1. Open LSB-embedded image in Photoshop
2. Add text using text tool
3. Save with "Quick Export as PNG"

Results

【Pixel Analysis】
   First changed pixel: Row 13, Column 43 (text area only)
   Changed pixels: 9.2% (text area only)

【LSB Message Area】
   Location: Row 0, Columns 0-37 (first 38 pixels)
   Status: Completely intact ✅

【Extraction Test】
   Original:        'Hello LSB!'
   After Photoshop: 'Hello LSB!'

   🎉 Decoding successful!

Why It Survived

Factor	Explanation
Export method	"Quick Export as PNG" has minimal optimization
Color profile	No conversion (embedded profile maintained)
Edit location	Text in center, LSB data in top-left → no overlap
File format	Saved as PNG (lossless)

Survival Conditions

flowchart TD
    A[Edit in Photoshop] --> B{Color profile conversion?}
    B -->|Yes| X[❌ Destroyed]
    B -->|No| C{Save format?}
    C -->|JPEG/WebP| X
    C -->|PNG| D{Edit location?}
    D -->|Top-left edited| X
    D -->|Center/bottom-right| E[✅ Survives]

Conclusion

LSB steganography is more durable than expected... but don't trust it

It's not simply "PNG re-save is OK" - where you edit matters. However, this is not a security feature - use only as an experimental watermark.

---

Future Plans

Text Decoration

• Custom font support
• Stroke (outline)
• Drop shadow
• Text rotation
• Background box

Advanced Steganography

• DCT (Discrete Cosine Transform) method
• DWT (Discrete Wavelet Transform) method
• Position shuffling with encryption key
• Error correction codes

Other

• Logo rotation & blend modes
• C2PA signature support

License

Apache License 2.0

Related Projects

• ComfyUI Master Guide
• comfyui-save-image-local - Original save node
• C2PA - Content Credentials standard

Contributing

Issues and Pull Requests are welcome.

Author

AICU Japan K. K.

• Web: https://aicu.jp
• GitHub: https://github.com/aicuai