Is it Stable Diffusion?
Check a suspected Stable Diffusion image: an AI score from the pixels, and whatever the file still says about where it came from.
Drop a photo to check it
or paste one with ⌘V / Ctrl+V
Have an API key?
With a key, checks here spend your credits instead of the free daily allowance. It stays in this browser.
How to detect Stable Diffusion images
Stable Diffusion images often carry their full recipe. AUTOMATIC1111 and Forge write the prompt, seed, sampler and model into a PNG text chunk called parameters, and ComfyUI stores the whole node graph. PhotoCheck reads both and returns the generator and a prompt excerpt. Converted to JPEG or uploaded to social media, the image has only the model score to go on.
What the pixels give away
- Anatomy errors in older models (SD 1.5, SD 2): extra fingers, mismatched eyes, limbs that fork.
- With SDXL and fine-tunes, an over-detailed look: every strand of hair crisp, every surface textured.
- Recognisable fine-tune styles, like glossy anime shading or a particular film-grain look applied everywhere.
- Strange object boundaries where a background bleeds into hair or clothing.
What the file may carry
PhotoCheck parses PNG tEXt, zTXt and iTXt chunks: A1111/Forge "parameters", ComfyUI "prompt" and "workflow", and InvokeAI's "invokeai_metadata". These are the most explicit provenance you will ever find on an AI image, and also the easiest to remove: one re-save as JPEG drops them.
How the score behaves
Stable Diffusion is open source and endlessly fine-tuned, so output varies more than any closed generator's. The score is useful but not uniform; a photoreal fine-tune can score lower than a stock SD image.
Stable Diffusion questions
Can I read the prompt of a Stable Diffusion image?
If the PNG still has its parameters or ComfyUI chunk, yes: the metadata check returns up to 300 characters of the prompt.
Does Stable Diffusion add an invisible watermark?
The reference scripts can, but many popular interfaces turn it off. PhotoCheck does not rely on it.
Other generators
- MidjourneyMidjourney, Inc.
- ChatGPT imagesOpenAI
- DALL·EOpenAI
- FLUXBlack Forest Labs
- Google Gemini and ImagenGoogle
- Adobe FireflyAdobe
- IdeogramIdeogram AI
- Leonardo.AiLeonardo Interactive (Canva)
- Grok (Aurora)xAI
- Meta AIMeta
- StyleGAN faces (This Person Does Not Exist)NVIDIA research, used by countless fake-profile sites
Checking images in bulk? The AI detection API runs the same check for one credit an image.