Stable Diffusion has no single, centrally-enforced metadata format — it depends entirely on which interface generated the image. The most common case, popularized by the AUTOMATIC1111 WebUI and still used by many forks, writes a PNG text chunk literally named parameters containing your full prompt, negative prompt, sampler, step count, CFG scale, seed, and model hash — essentially the complete recipe to reproduce the image, in plain text, with no C2PA manifest involved.
Unlike OpenAI or Adobe, there's no single company controlling how Stable Diffusion models get deployed — it runs through dozens of different UIs, each with its own metadata conventions:
tEXt chunk named parameters with the full generation recipe as plain text.DigitalSourceType fields voluntarily, following IPTC's 2024 guidance, but this isn't universal or guaranteed.Because the full prompt, seed, and model are stored as legible text, anyone who inspects the file can reproduce your exact output, see your negative prompt (sometimes revealing what you were trying to avoid, which can itself be telling), and in some workflows, see the specific fine-tuned model or LoRA you used — information creators often don't intend to share publicly alongside the image.
Because this is stored as an ordinary PNG metadata chunk — not a cryptographically signed manifest — it's straightforward to both inspect and remove. AiEraser's View mode will show you any text chunks present (our scanner flags any non-standard PNG chunk rather than only recognizing a fixed list, so interface-specific conventions like this don't get silently missed), and Remove mode strips it via canvas re-encoding.
This reflects the widely-documented community convention for AUTOMATIC1111-family WebUIs. Since Stable Diffusion runs through many independent interfaces, check your specific tool's documentation for its exact metadata behavior.