Stable Diffusion Metadata: PNG Text Chunks, Prompts, and How to Remove Them

Stable Diffusion is unique among major AI image generators in the sheer volume of metadata it embeds in its outputs. While tools like Midjourney embed a few identifying fields and DALL-E 3 uses C2PA credentials, Stable Diffusion — through its most popular interfaces, AUTOMATIC1111 and ComfyUI — embeds a comprehensive record of the entire generation process directly in the image file. This includes your full prompt, your negative prompt, the model checkpoint you used, the LoRA weights, the seed, the number of steps, the CFG scale, the sampler, and in ComfyUI's case, the complete workflow graph as a JSON object.

For privacy-conscious creators, this is a significant concern. The embedded data is not just an AI identifier — it is a detailed record of your creative process, your preferred models, your prompt engineering techniques, and potentially your artistic style. This guide explains exactly where this data lives in Stable Diffusion output files, how to read it, and how to remove it completely.

Where Stable Diffusion Stores Metadata

AUTOMATIC1111 (A1111)

AUTOMATIC1111 stores generation parameters in two locations depending on the output format. For PNG files, parameters are stored in a tEXt chunk with the key 'parameters'. This chunk contains a plain-text string with the positive prompt, negative prompt (prefixed with 'Negative prompt:'), and a comma-separated list of generation parameters (Steps, Sampler, CFG scale, Seed, Size, Model hash, Model, LoRA hashes, Version). For JPEG files, A1111 stores the same information in the EXIF UserComment field, encoded as UTF-8.

ComfyUI

ComfyUI takes metadata embedding to an extreme: it stores the complete workflow graph as a JSON object in a PNG tEXt chunk with the key 'workflow'. This JSON contains every node in your workflow, every connection between nodes, every parameter value, and the complete model configuration. For complex workflows with multiple models, ControlNet inputs, and custom nodes, this JSON can be several hundred kilobytes in size — and it is embedded in every output image by default.

ComfyUI also stores a 'prompt' chunk containing a simplified version of the workflow as a JSON object. Both the 'workflow' and 'prompt' chunks are present in ComfyUI outputs unless you explicitly configure the tool to suppress them.

Reading Stable Diffusion Metadata

Before removing metadata, it is useful to know how to read it — both to understand what is being removed and to verify that removal was successful. There are several ways to read Stable Diffusion metadata.

ExifTool with the command 'exiftool -PNG:parameters filename.png' will display the A1111 parameters chunk. For ComfyUI workflow data, 'exiftool -PNG:workflow filename.png' will output the workflow JSON. For a complete view of all PNG text chunks, 'exiftool -a -u -g1 filename.png' shows everything.

Alternatively, the AUTOMATIC1111 web interface has a built-in 'PNG Info' tab that reads and displays the parameters from any Stable Diffusion PNG. This is useful for verifying that a cleaned image no longer contains generation parameters.

Why Standard Metadata Removal Fails for Stable Diffusion

Many creators assume that saving a Stable Diffusion PNG as a JPEG will remove the metadata. This is partially true: JPEG does not support PNG tEXt chunks, so the 'parameters' and 'workflow' chunks are lost in the conversion. However, A1111 re-embeds the parameters in the EXIF UserComment field when saving as JPEG, so the metadata is not actually lost — it is just moved to a different location.

Similarly, opening a Stable Diffusion PNG in an image editor and re-saving it does not reliably remove the tEXt chunks. Many image editors preserve unknown chunks when re-saving PNG files to maintain compatibility. GIMP, for example, preserves tEXt chunks by default. Photoshop's behavior varies by version and export settings.

The only reliable way to remove Stable Diffusion metadata is to either explicitly delete the tEXt chunks using a tool that understands PNG chunk structure, or to re-encode the image from its pixel data using a method that does not carry over any metadata containers.

Complete Removal Methods

Method 1: Canvas Re-encoding (Recommended)

Drawing the image onto an HTML5 Canvas and exporting it creates a new PNG or JPEG file that contains only pixel data — no tEXt chunks, no EXIF, no metadata of any kind. This is the approach used by BlankAI, and it is the most reliable method because it is format-agnostic: it works identically for A1111 PNGs, ComfyUI PNGs, and JPEG outputs.

Method 2: ExifTool Chunk Deletion

ExifTool can delete specific PNG chunks with targeted commands. To remove A1111 parameters: 'exiftool -PNG:parameters= filename.png'. To remove ComfyUI workflow data: 'exiftool -PNG:workflow= -PNG:prompt= filename.png'. To remove all PNG text chunks and EXIF data: 'exiftool -all= -PNG:all= filename.png'. Note that the '-overwrite_original' flag will overwrite the original file; without it, ExifTool creates a backup.

Protecting Your Workflow and Prompt Engineering

Beyond the AI detection concern, there is a creative privacy dimension to Stable Diffusion metadata removal. Your prompts, model choices, LoRA configurations, and workflow designs represent significant creative and technical investment. When you share an image with embedded A1111 parameters, you are also sharing your prompt engineering techniques, your preferred models, and potentially proprietary workflow designs.

For creators who have developed distinctive styles through careful prompt engineering and model fine-tuning, metadata removal is not just about avoiding AI detection — it is about protecting intellectual property. The 'workflow' chunk in ComfyUI outputs is particularly sensitive: it contains the complete blueprint of your image generation process, which competitors or clients could use to replicate your style.