Stable Diffusion Metadata: PNG Text Chunks, Prompts and Workflows

Local Stable Diffusion interfaces are generous with metadata by design: they save everything needed to reproduce an image inside the image. That is convenient for you and revealing for anyone you share the file with.

AUTOMATIC1111 and Forge

Both write a PNG text chunk named “parameters”. It holds the prompt, the negative prompt, and a line with steps, sampler, CFG scale, seed, size, model hash and model name, plus any extras such as LoRAs and hires-fix settings. When saving JPEG or WebP, they put the same text in the EXIF UserComment field. A setting in AUTOMATIC1111 stops it from writing the PNG chunk.

ComfyUI

ComfyUI's SaveImage node writes two PNG chunks: “prompt”, the executed graph, and “workflow”, the full editable workflow. Dropping the PNG onto a ComfyUI canvas restores the whole graph, including node settings and model file names. Starting ComfyUI with --disable-metadata stops it.

Invisible watermarks in some pipelines

The original CompVis Stable Diffusion 1.x script, Stability's SDXL and SVD scripts, and the diffusers SDXL pipeline when the invisible-watermark package is installed all add a DWT-DCT watermark to the pixels by default. AUTOMATIC1111, Forge and ComfyUI do not. Metadata removal does not affect this watermark.

Data hidden in pixels

Some extensions store generation parameters in the lowest bits of the alpha channel instead of in text chunks. A lossless PNG copy keeps those pixel values, so it keeps that data. Saving as JPEG discards the alpha channel and the hidden bits with it.

How to remove Stable Diffusion metadata

1. 1Drop your renders into BlankAI. It lists parameters and workflow chunks by name.
2. 2Keep “Same as original” to get lossless PNGs with transparency.
3. 3Select Remove Metadata and save the clean copies.
4. 4Keep your originals: they are the easiest way to reload a workflow later.