Loading...
Loading...
10 step-by-step guides to extracting the prompt, seed, and settings from an AI-generated image — by tool, by file format, and when the metadata has been stripped. Each links the free metadata inspector so you can try it immediately.
10 guides
Reading one PNG at a time does not scale. Batch extraction reads the prompt chunks across a whole folder and turns them into a searchable index.
JPEG and WebP do not carry PNG text chunks, so a prompt saved as a PNG chunk is usually lost on conversion — though some tools write EXIF/XMP that may survive.
AI tools store prompts in PNG text chunks — tEXt (plain), zTXt (compressed), and iTXt (UTF-8). Reading them is how prompt recovery from PNGs actually works.
Platforms like Discord, X, Instagram, and Reddit re-encode images on upload, which strips embedded prompt metadata — the saved copy carries nothing back to the prompt.
A screenshot is a brand-new image captured by your operating system — it carries none of the original generation metadata, so the prompt is not inside it.
ComfyUI embeds the full workflow and prompt inside every output PNG as text chunks — as long as the file is the original, unconverted PNG, the prompt is fully recoverable.
Recent Midjourney downloads embed the full prompt in the file’s Description field; for older or shared images, the Job ID and your web history are the route back to the prompt.
Most Stable Diffusion interfaces write the full generation settings into a "parameters" PNG text chunk — prompt, negative prompt, sampler, steps, CFG, seed, and model.
The seed is stored in the same metadata as the prompt. If the file still has its generation metadata, the seed is readable; if it was stripped, the exact seed cannot be recovered from pixels.
Embedded metadata is fragile: uploads, conversions, and edits strip it. If you only hold a stripped copy, the prompt is not inside the pixels — recovery means finding the original or the tool’s history.