Understanding AI Metadata - Ktiseos-Nyx/Dataset-Tools GitHub Wiki

A Beginner's Guide to AI Image Metadata

The core purpose of Dataset-Tools is to help you see and understand the data hidden inside your AI-generated images. This "metadata" is the digital recipe of your creation, and understanding it is key to improving your workflow.

What is AI Metadata?

When an AI image generator creates an image, it often embeds the exact settings used into the image file itself (most commonly a PNG) or saves them in a matching text file. This data is the key to understanding how an image was made.

This metadata can include:

  • The prompt and negative prompt.
  • Generation parameters like the seed, steps, and sampler.
  • Information about the models, LoRAs, and other resources used.
  • Complex workflow data from tools like ComfyUI.

Dataset-Tools finds all of this information, parses it, and displays it in a clean, organized way.

Glossary of Common Terms

Here are some of the most common parameters you will see in the metadata panel:

  • Prompt: The primary text description used to generate the image.
  • Negative Prompt: A list of terms used to tell the AI what you want to avoid in the image.
  • Seed: The starting number for the random generation process. Using the same seed and settings will produce an almost identical image. A seed of -1 means a random seed was used.
  • Steps: The number of iterations the sampler runs. More steps can add more detail but take longer.
  • Sampler: The algorithm used to create the image (e.g., Euler a, DPM++ 2M Karras). Different samplers can have a significant effect on the final style.
  • CFG Scale (Classifier Free Guidance): A value that controls how strictly the AI must follow your prompt. Lower values give the AI more creative freedom, while higher values force it to adhere more closely.
  • Size: The width and height of the image in pixels (e.g., 1024x1024).
  • Model / Model Hash: The main checkpoint model used for generation. The "hash" is a unique ID for that specific model file, which is useful for identifying it.
  • LoRA / LoCon: Low-Rank Adaptation models used to add specific styles, characters, or concepts.
  • VAE (Variational Autoencoder): The component that translates the AI's internal data into the final pixel image. A different VAE can affect colors and fine details.
  • Upscaler / Hires. Fix: The method used to increase the image's resolution, often with its own specific settings like "Denoising strength."

How Dataset-Tools Helps

Our tool is specifically designed to find and neatly display this information from various sources:

  • Image Files (PNG, JPG, WebP): Dataset-Tools reads the metadata directly from the image file and presents it in a human-readable format. For ComfyUI, it can even parse the complex workflow JSON embedded within the PNG.
  • Model Files (.safetensors): For files like LoRAs, the tool can read internal metadata, which might include training information, recommended trigger words, and the base model used.
  • Text Files (.txt, .caption): If a text file shares the same name as an image (e.g., car.png and car.txt), the tool will treat its contents as the image's caption or metadata.

Why This Metadata Matters

Understanding this data unlocks several powerful capabilities in your creative workflow:

  • Reproducibility: Copy the exact settings to create a similar image again.
  • Organization: Group, search, or filter your datasets based on the models or prompts used.
  • Learning: Study the parameters used by other artists to understand how they achieve certain styles.
  • Troubleshooting: Analyze the settings to figure out why a generation didn't turn out as expected.
  • Sharing: Provide others with the exact "recipe" for your image.