PNG Info
Reads the generation data stored inside an AI image: the full prompt, the negative prompt, the seed, the sampler and the model name from Stable Diffusion, ComfyUI or NovelAI. PNG, JPG and WebP are read, and the file is checked for labels that mark it as AI-generated.
This tool does not upload your files. Close the tab and nothing is left behind.
Recovering the prompt from a generated image
This tool reads the generation data that AI image tools leave inside their output. A PNG from Stable Diffusion WebUI (A1111) carries a `parameters` text chunk, which is split into Prompt, Negative prompt, and fields such as Steps, Sampler, CFG scale, Seed and Model; prompts in Korean, Japanese or Chinese land in an iTXt chunk instead, and that is read too. A JPG or WebP saved by A1111 holds the same string in the EXIF UserComment field, ComfyUI's node graph is reduced to its prompt texts and sampler settings, and NovelAI's JSON comment is read as well. Other text chunks are shown as they are.
Above the fields sits a verdict on how the file was made, the same one the EXIF Viewer shows: labelled as AI-generated when the file carries generation settings, an IPTC digital source type, a Content Credentials (C2PA) manifest or a known generator in its software field, a camera record when it has one and no label, and nothing when the file is silent. Each parsed field has its own copy button. Other text rows are displayed without individual copy buttons, and PNG header values such as dimensions and bit depth are listed separately.
Some services never write the prompt in the first place: images from DALL-E or recent Midjourney carry no prompt to read (older Midjourney PNGs keep one in the Description field, and it is shown), though they may still carry an AI label that the verdict reports. Nothing here analyses the pixels, so a file with its metadata removed shows no label, and that is not evidence of a photograph. The reverse is worth knowing as well, because sharing your own images means shipping your prompts to whoever downloads them; running them through the Remove EXIF tool deletes the text chunks and the comment field with no loss of quality. Either way the file is read inside your browser and is never sent anywhere.
How to use
- 1
Drop in a PNG, JPG or WebP file.
- 2
Any generation data is split into labelled fields, and the card above says whether the file carries an AI label.
- 3
Use the copy button next to a field to take just the prompt.
Frequently asked questions
- Which images carry this?
- Original files from Stable Diffusion WebUI (A1111), ComfyUI, NovelAI and similar tools. A PNG holds the settings in a text chunk, including the iTXt chunk used for prompts in Korean, Japanese or Chinese; a JPG or WebP saved by A1111 holds the same string in the EXIF UserComment field.
- Why do I get nothing?
- The data is lost when the file is re-saved by an editor, sent through a messenger, or downloaded back from a site that strips it. Some services, such as DALL-E and recent Midjourney, never write the prompt into the file at all, although they may still add an AI label that the card reports. Older Midjourney PNGs keep the prompt and Job ID in the Description field, and those are read.
- How does it decide a file is AI-made?
- By what the file itself declares: generation settings, an IPTC digital source type, a Content Credentials (C2PA) manifest or a known generator in the software field. It does not analyse the pixels, so an image with its metadata removed shows no label, and that is not proof it is a photograph.
- How do I stop my own images from carrying the prompt?
- Run them through the Remove EXIF tool, which deletes the text chunks and the comment field without any loss of quality, before you share them.