Custom nodes for ComfyUI that extract text prompts from PNG image metadata. Recover prompts from any ComfyUI-generated image without needing the original workflow.

✨ Features

Extract prompts from PNG metadata — Recovers positive and negative prompts from ComfyUI images
Smart prompt classification — Automatically identifies positive vs negative prompts
Multiple extraction modes — Longest string, positive only, negative only, or all prompts
Three node variants — Single image, file path, and batch processing
Full metadata access — Outputs complete workflow JSON for advanced use cases
Zero external dependencies — Uses only PIL and standard Python libraries

📦 Installation

Method 1: ComfyUI Manager (Recommended)

Search for "DJZ Prompt Extractor" in ComfyUI Manager and click Install.

Method 2: Manual Installation

Clone this repository into your ComfyUI custom nodes directory:

cd ComfyUI/custom_nodes
git clone https://github.com/MushroomFleet/ComfyUI_PromptExtractor_nodes.git

Restart ComfyUI after installation.

🎛️ Nodes

DJZ Prompt Extractor

The main extraction node. Accepts an image input with an optional file path for metadata access.

Inputs: | Input | Type | Description | |-------|------|-------------| | image | IMAGE | ComfyUI image tensor | | image_path | STRING | Optional: Direct path to PNG file (recommended for full metadata) | | extraction_mode | ENUM | How to extract prompts |

Outputs: | Output | Type | Description | |--------|------|-------------| | prompt | STRING | Extracted positive prompt | | negative_prompt | STRING | Extracted negative prompt | | metadata_json | STRING | Full metadata as JSON |

DJZ Prompt Extractor (Path)

Simplified node that takes only a file path. Use when you have the path to a PNG file.

Inputs: | Input | Type | Description | |-------|------|-------------| | image_path | STRING | Path to PNG file | | extraction_mode | ENUM | How to extract prompts |

Outputs: Same as main node.

DJZ Prompt Extractor (Batch)

Processes all PNG files in a directory. Useful for bulk prompt extraction.

Inputs: | Input | Type | Description | |-------|------|-------------| | directory_path | STRING | Path to folder containing PNGs | | max_images | INT | Maximum images to process (1-100) | | extraction_mode | ENUM | How to extract prompts |

Outputs: | Output | Type | Description | |--------|------|-------------| | prompts_list | STRING | All extracted prompts (separated by ---) | | filenames_list | STRING | Corresponding filenames |

🔧 Extraction Modes

| Mode | Description | |------|-------------| | longest_string | Returns the longest string found in metadata (default) | | positive_only | Attempts to identify and return only the positive prompt | | negative_only | Attempts to identify and return only the negative prompt | | all_prompts | Returns all found prompts concatenated with | separator |

💡 Usage Examples

Basic Prompt Recovery

Connect a Load Image node to the extractor to recover its original prompt:

[Load Image] → [DJZ Prompt Extractor] → [Show Text]
                                      → [CLIP Text Encode]

Re-generate with Extracted Prompt

Use the extracted prompt directly in a new generation:

[Load Image] → [DJZ Prompt Extractor] 
                    ↓ prompt
               [CLIP Text Encode (Positive)]
                    ↓ negative_prompt  
               [CLIP Text Encode (Negative)]
                    ↓
               [KSampler] → [Save Image]

Batch Extract from Output Folder

Extract prompts from all images in your output directory:

[DJZ Prompt Extractor (Batch)]
    directory_path: "ComfyUI/output"
    max_images: 50
         ↓
    [Show Text] — View all extracted prompts

Using File Path for Full Metadata

For complete metadata extraction, provide the file path directly:

[Load Image]
    ↓ image
[DJZ Prompt Extractor]
    image_path: "/path/to/your/image.png"
         ↓
    prompt, negative_prompt, metadata_json

⚠️ Important Notes

PNG Metadata Preservation

ComfyUI embeds workflow and prompt data in PNG tEXt chunks. This metadata is lost when:

Images are converted to tensors (IMAGE type in ComfyUI)
Images are re-saved by other applications
Images are uploaded to social media or image hosts
Images are converted to JPEG or other formats

For best results: Use the image_path input with the original PNG file path, or use the Path variant node.

Prompt Classification

The node uses two methods to classify prompts as positive or negative:

Content analysis — Looks for common negative prompt keywords (ugly, blurry, watermark, deformed, etc.)
Workflow tracing — Analyzes node connections to determine if a CLIPTextEncode connects to a sampler's positive or negative input

Classification may not be 100% accurate for unusual workflows or unconventional negative prompts.

🔍 How It Works

Read PNG chunks — Parses the PNG file structure to find tEXt and iTXt metadata chunks
Extract JSON — ComfyUI stores workflow and prompt data as JSON strings
Find CLIPTextEncode nodes — Locates all text encoding nodes in the workflow
Extract text inputs — Pulls the text field from each encoder's inputs
Classify prompts — Determines positive vs negative based on content and connections
Return results — Outputs the classified prompts and full metadata

🌐 Related Projects

This node is the Python/ComfyUI implementation of the prompt extraction logic. A browser-based JSX version is also available:

ComfyUI-Prompt-Extractor-JSX — React component for extracting prompts in web applications

🐛 Troubleshooting

"No prompts extracted" or empty output

Ensure you're using the original PNG from ComfyUI's output folder
Check that the file hasn't been re-saved or converted
Try providing the image_path directly instead of relying on the tensor

Prompts are misclassified

Use all_prompts mode to see everything found
Check the metadata_json output for the raw workflow data
Some custom workflows may not follow standard positive/negative patterns

File not found errors

Use absolute paths when possible
Check that the path doesn't contain special characters
Verify the file exists and is readable

📄 License

MIT License — free for personal and commercial use.

📚 Citation

Academic Citation

If you use this codebase in your research or project, please cite:

@software{comfyui_prompt_extractor_nodes,
  title = {ComfyUI Prompt Extractor Nodes: Custom nodes for extracting text prompts from PNG metadata},
  author = {Drift Johnson},
  year = {2025},
  url = {https://github.com/MushroomFleet/ComfyUI_PromptExtractor_nodes},
  version = {1.0.0}
}

Donate:

Support This Project

If you found this useful, please star the repo — it helps others discover it!