What Stable Diffusion files can record
Stable Diffusion is an open model family that people run through many different programs, so there is no single file signature for it. Some of those programs write their settings into the image. The AUTOMATIC1111 web UI, one of the most widely used front ends, adds the generation parameters to PNG files as a text chunk, usually called PNG Info: the prompt, the negative prompt, and fields such as Steps, Sampler, CFG scale, Seed and the model name. Anyone with a PNG chunk viewer, or the PNG Info tab of the web UI itself, can read it back.
Source: AUTOMATIC1111 web UI features: PNG info (opens in a new tab)
What our checker reads
The free Image Evidence checker looks for an AUTOMATIC1111-style parameter block inside the PNG text chunks of the file you choose, and reports it as a declared generation signal when at least two of the standard fields are present. It also reads IPTC and XMP digital source type declarations and any C2PA Content Credentials. The file is read inside your browser tab and is not uploaded. The checker does not display or keep the prompt or parameter text; it tells you that the structure exists, which is the part that matters for an origin question.
Other tools write different fields
ComfyUI can embed its whole node workflow in saved PNG images so the file can be reopened to recreate the setup, and other interfaces and hosted services write their own fields or none at all. Our checker does not currently treat a ComfyUI workflow as a generation signal, so open such files in ComfyUI or a metadata viewer if you need to see the workflow. Because Stable Diffusion runs on people's own hardware, scripts can be changed freely: a missing parameter block or a missing watermark says very little about where a file started.
Source: ComfyUI examples: workflow metadata in images (opens in a new tab)
Why the block disappears
PNG text chunks survive only as long as the file stays a PNG and nobody removes them. Converting to JPEG or WebP, taking a screenshot, cropping in many editors, uploading to a social platform or running a metadata cleaner usually drops them. The block is also plain, unsigned text: it can be deleted from a generated image or pasted into a photograph. Treat a found block as a strong lead about the file's history, and an absent block as no evidence either way.
Read a parameter block carefully
A parameter block shows that the file passed through a Stable Diffusion style tool. It does not say whether the whole scene was generated, because image-to-image and inpainting workflows start from an existing picture and write the same kind of block. If the image matters, ask the creator for the source image and the workflow, compare the claimed model and settings with the result, and keep a note of what you verified and what remains unknown. Image Evidence has not published a Stable Diffusion detection rate.
How do I check if an image was made with Stable Diffusion?
Start with the original PNG if you can get it. Open it in the checker above: if an AUTOMATIC1111-style parameter block is present, the result shows a declared generation signal. A PNG chunk viewer or the web UI's PNG Info tab shows the full text. If the file has been converted, screenshotted or downloaded from social media, the block is probably gone, and you will need other evidence such as the source account or the creator's working files.
Can I see the prompt used to make an image?
Only when the generator saved it and the copy you have still contains it. AUTOMATIC1111 and similar tools store the prompt in PNG Info by default, and ComfyUI stores a workflow. Image-to-prompt tools that guess a prompt from the pixels are writing a new description, not reading the original. Our checker deliberately does not show prompt text, because prompts can contain private or sensitive material.
Can AI detectors detect Stable Diffusion images?
Detectors trained on Stable Diffusion outputs often flag unedited images from the versions they were tested on. Fine-tuned models, LoRAs, upscaling, editing and compression make results less predictable, and research has shown that detectors can generalize poorly to generators missing from their training data. A score should be read alongside the file's metadata and the claim attached to the image.