How to Inspect and Read Hidden AI Image Metadata
When generative AI models create an image, they write technical data directly into the file header. This guide explains how to inspect and interpret hidden AI metadata in PNG and JPEG images.
What Information is Hidden Inside AI Images?
Depending on the generator software used, image files contain structured key-value parameters:
- Text Prompt: The literal instructions provided to the text-to-image model.
- Negative Prompt: Concepts and attributes instructed to be excluded (common in Stable Diffusion).
- Seed Number: The exact random noise initialization integer used to produce the visual composition.
- Sampler & CFG Scale: The algorithm (e.g. DPM++ 2M Karras, Euler a) and Classifier-Free Guidance strength.
- Model Checkpoint Hash: Unique identifier specifying the exact neural network weights used.
Frequently Asked Questions
How do I see what prompt was used to make an AI image?
Drop the image into the viewer. If it came from Stable Diffusion, Forge or ComfyUI, the full prompt is stored in a PNG text chunk and will be shown along with the negative prompt, seed, sampler, CFG scale and model. Images from DALL-E 3 or Firefly do not store the prompt at all.
What is PNG Info?
PNG Info is the name Stable Diffusion interfaces use for the generation settings written into a PNG file. It is a plain-text block under the key "parameters" holding the prompt, negative prompt, steps, sampler, CFG scale, seed and model hash. This viewer reads the same block without needing Stable Diffusion installed.
Which AI generators store the prompt inside the file?
Local Stable Diffusion tools (AUTOMATIC1111, Forge, ComfyUI, InvokeAI) store the full prompt and every setting. Midjourney can carry the prompt in EXIF or XMP depending on where you downloaded it. DALL-E 3, ChatGPT images and Adobe Firefly store a marker identifying the image as AI-generated but never the prompt text.
Why does my image show no prompt?
Almost always because the file was re-encoded after generation. Social platforms strip metadata on upload, screenshots discard it entirely, and converting PNG to JPEG or WebP drops the text chunks. Always test the original download rather than a copy that has passed through another service.
Can I recover a ComfyUI workflow from a PNG?
Yes. ComfyUI writes two JSON graphs into the file, one executable and one describing the visual node layout. The viewer surfaces the conditioning text, checkpoint and sampler settings as readable fields instead of raw JSON, and the original PNG can still be dragged onto a ComfyUI canvas to rebuild the graph.
What do seed, CFG scale and sampler actually mean?
Seed is the integer that initialises the noise, so the same seed with identical settings reproduces almost the same image. CFG scale controls how strictly the model follows the prompt. Sampler is the denoising algorithm, and different samplers give visibly different results at the same seed.
Is it OK to look at the prompt on someone else's image?
Reading metadata a publisher chose to leave in a public file is exactly what the format allows, and studying a prompt to learn a technique is normal practice. Passing someone's prompt off as your own, or scraping an artist's catalogue to clone their style commercially, is where it stops being reasonable.
Does my image get uploaded anywhere?
No. The file is decoded in your browser and never leaves your device, which matters because prompts routinely contain unreleased product names and client briefs. There is no account, no upload and no server-side copy.