When you generate an image using ComfyUI or AUTOMATIC1111's Stable Diffusion WebUI, you are not just saving a picture. You are saving a full blueprint of how that image was created.
Share that PNG on an art forum or Discord channel, and anyone who downloads it can drag and drop it into their own UI to recreate your workflow. They get your exact positive and negative prompts, seed numbers, sampler steps, custom node groupings, and model hashes. If you want to keep your techniques private, that is a problem.
Here is where AI image generators hide this data, why standard image editors can ruin your file when you try to remove it, and how to strip your AI metadata locally without losing a single pixel of quality.
The Anatomy of an AI-Generated PNG
The PNG format uses a sequential, chunk-based layout. After an 8-byte signature that verifies the file, all data is distributed into structured binary chunks.
The visual pixels live inside IDAT chunks, but the format also allows optional ancillary chunks that can store almost anything. Generative AI frameworks use these text chunks to embed their data:
- tEXt: Standard uncompressed text strings.
- zTXt: Compressed text parameters using the zlib deflate algorithm.
- iTXt: International UTF-8 strings, which are often used for raw XML-based datasets.
How ComfyUI Stores Workflows
ComfyUI embeds the complete parameters of its generative nodes inside PNG text chunks under two primary keywords:
- workflow: A massive JSON representation of your visual user interface. This lists your exact coordinate maps, visual link arrays, custom node groupings, and even the color coding of your nodes.
- prompt: A structured serialization of the execution graph sent to the generation backend. This includes sampler parameters (seed, steps, cfg, scheduler), model loaders (identifying exact checkpoints, LoRAs, and ControlNets), and the raw text prompt values.
Because the graph data can be large, ComfyUI often compresses it inside a zTXt chunk to keep the PNG file size down.
The AUTOMATIC1111 Approach
In contrast to ComfyUI's complex multi-chunk model, the AUTOMATIC1111 WebUI leverages a singular key-value store. It writes a structured string into a tEXt chunk or an EXIF block under the keyword parameters.
This string contains your positive prompt, your negative prompt, and is capped by a comma-delimited block of generation settings (Steps, Sampler, CFG, Seed, Model hash). If processed by standard digital asset managers, this block is frequently parsed and duplicated into standard IPTC or XMP fields, scattering your prompt across multiple metadata domains.
The HTML5 Canvas Trap: Why Standard Tools Ruin Quality
Many web-based tools and generic scripts try to clean images by loading the file into an HTML5 Image element, drawing it onto a <canvas>, and calling toBlob() to generate a new output.
This does strip the AI metadata, but it also re-encodes the image. Chromium enforces 4:2:0 chroma subsampling when you export at anything under 100% quality, blurring fine details and color boundaries. Drawing wide-gamut graphics onto a standard canvas also truncates colors down to sRGB.
To remove AI workflow data without touching the pixels, you need a tool that performs binary chunk manipulation - reading the raw byte array, dropping only the tEXt, zTXt, and iTXt chunks, and leaving the IDAT image blocks untouched. The resulting file is bit-for-bit identical in its visual data.
The Hard Truth About AI Detection
It is critical to understand the limits of metadata removal. Using a Metadata Remover to strip ComfyUI nodes, EXIF data, or AUTOMATIC1111 parameters only removes self-declared structural markers.
Removing metadata does NOT defeat pixel-based AI detectors.
Statistical classifiers do not read your PNG text chunks. They analyze texture statistics, high-frequency noise, upsampling artifacts, and structural relationships between neighboring pixels. Deep invisible watermarking systems (like Google DeepMind's SynthID) work differently still - they inject patterns directly into the pixel values. These patterns survive metadata stripping, format conversions, and lossy compression.
Stripping metadata protects your workflow IP and prompt secrets. It does not make an AI-generated image undetectable to a trained pixel classifier.
To clean your PNGs safely, drop your files into our Metadata Remover. It runs 100% in your browser, strips the AI chunks at the binary level, and never uploads your files to a server.