Auto-Extraction - Detects sampler params, prompts, model/VAE, and LoRAs directly from your workflow
Performance Metrics - Auto-tracks GPU usage, VRAM peak, generation time, and software versions
Multi-Format - PNG, JPEG, and WebP with metadata injection
Filename Patterns - Placeholder-based filenames with sanitization
A1111/Civitai Compatible - Saves metadata in tEXt chunk ("parameters") recognized by Automatic1111, Civitai, and most SD tools
Image MetaHub Compatible - Saves extended metadata in iTXt chunk ("imagemetahub_data") with full workflow JSON
Model Hashes - Calculates SHA256 hashes (AutoV2 format) for models and LoRAs
IMH Pro Fields - Support for user tags, notes, and project names
Performance - Hash caching and graceful degradation ensure fast generation
Never Fails - Silent fallback on errors - your generation never stops

Installation

Method 1: Clone Repository (Recommended)

Navigate to your ComfyUI custom nodes directory:
```
cd ComfyUI/custom_nodes
```

Clone this repository:

git clone https://github.com/LuqP2/ImageMetaHub-ComfyUI-Save.git

Install dependencies:

cd ImageMetaHub-ComfyUI-Save
pip install -r requirements.txt

Restart ComfyUI

Method 2: Manual Installation

Download this repository as ZIP
Extract to ComfyUI/custom_nodes/ImageMetaHub-ComfyUI-Save
Install dependencies: pip install Pillow>=10.0.0 numpy>=1.24.0 piexif>=1.1.3
Restart ComfyUI

Usage

Quick Start

Add "MetaHub Save Image" node to your workflow
Connect the images output from your decoder/sampler
Generate! The node auto-detects and saves metadata.

Auto-detected fields:

seed, steps, cfg
sampler and scheduler
model name + hash
positive/negative prompts
LoRAs + weights
VAE name

Override (Optional)

Generation Parameters: Connect any override input to replace auto-detected values:

seed, steps, cfg
sampler_name, scheduler
model_name, vae_name
positive, negative
denoise

Performance Metrics (Advanced): For custom benchmarking workflows:

vram_peak_mb - Override VRAM peak measurement
gpu_device_override - Override GPU device name
generation_time_override - Override generation time tracking

All overrides use forceInput: True (hidden from UI, connection-only).

Filename Pattern

Use filename_pattern to customize names (default ComfyUI_%counter%). Invalid filename characters are sanitized and missing values become unknown. filename_prefix is deprecated but still supported.

Placeholder	Description	Example
`%counter%`	Auto-increment (00001, 00002...)	00042
`%seed%`	Seed used	1234567890
`%date%`	Date YYYY-MM-DD	2025-01-15
`%time%`	Time HH-MM-SS	14-30-22
`%datetime%`	Date + time	2025-01-15_14-30-22
`%model%`	Model name (no extension)	dreamshaper_v8
`%sampler%`	Sampler name	euler_ancestral
`%steps%`	Steps	20
`%cfg%`	CFG scale	7.5
`%width%`	Image width	1024
`%height%`	Image height	768

File Formats

file_format: PNG, JPEG, or WebP
quality: JPEG/WebP quality (1-100)
output_path: custom output directory (empty = ComfyUI default)
PNG uses tEXt "parameters" and iTXt "imagemetahub_data"
JPEG/WebP use EXIF UserComment (A1111) and ImageDescription (IMH JSON)

If structured metadata cannot be written, the image is still saved.

Multi-Sampler Heuristic

When multiple samplers exist, the node prefers the sampler that feeds the latent used by the VAEDecode connected to the saved images; it falls back to the first sampler found.

IMH Pro Fields (Optional)

Organize your generations with Image MetaHub Pro features:

user_tags: Tag your images (e.g., "portrait, fantasy, character")
notes: Add generation notes or experiments
project_name: Group images by project

Performance Metrics (Auto-Tracked)

The node automatically collects GPU and performance metrics for every generation:

Tier 1 - Critical Metrics:

VRAM Peak - Peak VRAM usage in MB (CUDA GPUs only)
GPU Device - GPU name (e.g., "NVIDIA GeForce RTX 3090")
Generation Time - Total generation time in milliseconds

Tier 2 - Very Useful:

Steps per Second - Performance benchmark (steps / generation_time)
ComfyUI Version - Your ComfyUI version

Tier 3 - Nice-to-Have:

PyTorch Version - PyTorch version with CUDA info
Python Version - Python runtime version

Auto-Detection:

✅ CUDA (NVIDIA) - Full metrics including VRAM tracking
✅ MPS (Apple Metal) - GPU device name (VRAM not available)
✅ CPU - Marks as "CPU (CUDA not available)"

Advanced Overrides (Optional): For custom benchmarking workflows, connect these inputs:

vram_peak_mb - Override VRAM peak (FLOAT)
gpu_device_override - Override GPU device name (STRING)
generation_time_override - Override generation time in seconds (FLOAT)

All metrics are stored in the analytics field of the Image MetaHub metadata chunk and displayed in the Performance section of Image MetaHub App.

Configuration (Optional)

If your models are in non-standard locations, set environment variables:

# Windows (PowerShell)
$env:COMFYUI_CHECKPOINT_PATH="D:\MyModels\Checkpoints"
$env:COMFYUI_LORA_PATH="D:\MyModels\Loras"
$env:COMFYUI_VAE_PATH="D:\MyModels\VAE"

# Linux/macOS
export COMFYUI_CHECKPOINT_PATH="/path/to/checkpoints"
export COMFYUI_LORA_PATH="/path/to/loras"
export COMFYUI_VAE_PATH="/path/to/vae"

Metadata Formats

A1111/Civitai Format (PNG: tEXt "parameters", JPEG/WebP: EXIF UserComment)

masterpiece, best quality, 1girl, portrait
Negative prompt: ugly, blurry
Steps: 20, Sampler: euler, CFG scale: 7.0, Seed: 12345, Size: 512x768, Model: mymodel, Model hash: abc1234567, Lora hashes: "detail: def9876543"

Image MetaHub Format (PNG: iTXt "imagemetahub_data", JPEG/WebP: EXIF ImageDescription)

{
  "generator": "ComfyUI",
  "prompt": "masterpiece, best quality, 1girl, portrait",
  "negativePrompt": "ugly, blurry",
  "seed": 12345,
  "steps": 20,
  "cfg": 7.0,
  "sampler_name": "euler",
  "model": "mymodel.safetensors",
  "model_hash": "abc1234567",
  "loras": [{"name": "detail.safetensors", "weight": 0.8}],
  "imh_pro": {
    "user_tags": "portrait, fantasy",
    "notes": "Experimental composition",
    "project_name": "Character Design"
  },
  "analytics": {
    "vram_peak_mb": 8234.5,
    "gpu_device": "NVIDIA GeForce RTX 3090",
    "generation_time_ms": 1523,
    "steps_per_second": 13.2,
    "comfyui_version": "0.1.0",
    "torch_version": "2.0.1+cu118",
    "python_version": "3.10.12"
  },
  "workflow": { }
}

Integration with Image MetaHub

This node is the official companion for Image MetaHub.

Images saved with this node are fully compatible with Image MetaHub's:

Instant parsing - MetaHub chunk extraction (no graph traversal needed)
Performance section - Displays GPU metrics, VRAM, generation time, and versions
Search and filtering - Full metadata indexing
LoRA detection - Automatic LoRA weight extraction
Analytics dashboard - Performance benchmarking
IMH Pro features - Tags, notes, and project organization

The Performance section in Image MetaHub App displays:

Generation time with smart formatting (ms/s/m+s)
VRAM usage with GPU percentage (e.g., "8.0 GB / 24 GB (33%)")
GPU device name
Steps per second benchmark
Software versions (ComfyUI, PyTorch, Python)

Troubleshooting

"Model hash is 0000000000"

Your model file wasn't found in standard paths
Set COMFYUI_CHECKPOINT_PATH environment variable
Or ensure models are in ComfyUI/models/checkpoints/

"LoRAs not detected"

Make sure you're using standard LoRA Loader nodes
Check that LoRAs are connected in the workflow
Custom LoRA nodes might not be detected

"Cannot create output directory"

Check folder permissions
Try using default output path (leave output_path empty)

"Performance metrics not showing in Image MetaHub"

Ensure you're using the latest version of both the node and Image MetaHub App
Performance section only appears for images saved with MetaHub Save Node
Legacy images saved with default ComfyUI Save Image won't have performance data

"VRAM peak is None/null"

VRAM tracking requires CUDA (NVIDIA GPUs)
MPS (Apple Metal) and CPU modes don't support VRAM tracking
This is expected behavior and not an error

Performance

First image: ~200ms overhead (hash calculation)
Subsequent images: ~60ms overhead (hashes cached)
Batch processing: Minimal overhead per image
GPU metrics: <1ms overhead (PyTorch API calls)
Version detection: <5ms overhead (cached after first call)

Hash calculation is cached in memory, so repeated generations with the same models are very fast. Performance metrics collection has negligible impact on generation speed.

License

MIT License

Contributing

Issues and pull requests welcome at https://github.com/LuqP2/ImageMetaHub-ComfyUI-Save

Credits

Developed for the Image MetaHub ecosystem.