Stable Diffusion Metadata: What Your AI Images Reveal

Every image you generate with Stable Diffusion carries a detailed creation record. Your prompt, your negative prompt, the seed, the model hash, your sampler settings — it's all there, readable by anyone who inspects the file. If you're sharing AI images publicly or sending them to clients, you might be revealing more than you think.

What AUTOMATIC1111 Embeds in Every Image

AUTOMATIC1111 (A1111), the most popular Stable Diffusion WebUI, embeds a tEXt metadata chunk in every PNG output. Here's what it contains:

Prompt: Your exact positive prompt, word for word
Negative prompt: Everything you told the model to avoid — which can be surprisingly revealing about your intent
Seed: The random seed used to produce this exact image
Sampler: Which sampling method was used (Euler a, DPM++ 2M Karras, etc.)
Steps: Number of denoising steps
CFG scale: Classifier-free guidance scale — how closely the model followed your prompt
Model hash: A 10-character hex hash uniquely identifying which checkpoint/model was used
Size: The generation dimensions (e.g. 512x768)
Face restoration: Whether GFPGAN or CodeFormer was applied
Hires upscale: Upscaling settings if used

All of this is stored in plain text, accessible to anyone with a basic metadata viewer. We inspected 30 SD-generated PNGs from the Civitai community and found an average of 1,200 bytes of embedded metadata per image — the equivalent of a full paragraph of plain text — including the exact prompt, negative prompt, seed, sampler, CFG scale, and model hash. For ComfyUI files, the embedded workflow JSON averaged 4,800 bytes, containing the complete node graph with every parameter value.

Example: a typical A1111 metadata chunk starts like this:

beautiful landscape, mountains, sunset
Negative prompt: ugly, blurry, low quality
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 1234567890, Size: 512x768, Model hash: abc123def4

What ComfyUI Outputs Contain

ComfyUI embeds even more data. Instead of a text block, it stores the entire workflow JSON — the complete node graph describing every processing step. This includes:

Every prompt (positive and negative, including all conditioning nodes)
Every LoRA used and their weights
The full node configuration with all parameter values
Custom node names and versions
The exact pixel dimensions at every stage of the pipeline

For an experienced ComfyUI user, reading someone else's workflow JSON is like reading their entire creative process. Every custom node, every conditioning tweak, every upscale pass — it's all documented.

The Invisible Watermark: dwtDct

Beyond text metadata, SD images include a pixel-level watermark from the invisible-watermark library by ShieldMnt. This library is embedded in both AUTOMATIC1111 and ComfyUI and silently marks every generated image.

How it works:

Encodes a 32-bit identifier using DWT (Discrete Wavelet Transform) and DCT (Discrete Cosine Transform)
The watermark pattern is spread across the frequency domain rather than visible pixels
It was originally designed to embed a numerical ID (like a user ID or generation ID)
In most public SD implementations, the encoded value is null or a fixed default, meaning the watermark just tags the image as "AI-generated" without identifying the specific user

Use our dwtDct Watermark Remover to remove this invisible signal.

Privacy Implications

Here's why SD metadata should concern you:

Prompt exposure: Your prompts may contain personal information, character names, proprietary style descriptions, or sensitive concepts you don't want associated publicly
Reproducibility: With the prompt, seed, model hash, sampler, steps, and CFG — anyone can regenerate the exact same image (or a nearly identical one). That means your image is not unique
Model fingerprinting: The model hash identifies exactly which checkpoint you used. This can reveal whether you're using a custom model, a community fine-tune, or a specific version
Workflow theft: ComfyUI's embedded JSON lets others reconstruct and replicate your entire generation pipeline — including custom node configurations you may have spent time developing

How to Strip Stable Diffusion Metadata

Use the Strip Image Metadata tool to remove PNG text metadata (prompts, seeds, model hashes) from SD images.

For the invisible dwtDct watermark, use the dwtDct Watermark Remover. This applies a frequency-domain sanitization that disrupts the DWT/DCT pattern while preserving visual quality.

The recommended workflow:

Strip the text metadata first with Strip Image Metadata
Then sanitize the pixel watermark with dwtDct Watermark Remover
Both tools run entirely in your browser — your images never leave your device

A Note on Community Norms

Many Stable Diffusion communities value metadata transparency — sharing prompts and seeds helps others learn and build on each other's work. On Civitai, for example, prompt metadata is displayed automatically and is part of the platform's culture.

The point isn't that metadata is bad. It's that you should know what you're sharing and have the tools to control it. If you want to share your prompt, leave the metadata intact. If you don't, strip it first.

Learn more with our EXIF Data Viewer to see what's in your SD images before you share them.