AI image detector
Stable Diffusion detector
Local Stable Diffusion tools write the whole recipe — prompt, negative prompt, sampler, seed, model — into the image file. Drop a PNG and GPTTrace will show it to you, or fall back to a pixel check if it has been stripped.
Free · no sign-up · your file never leaves your device
Drop a Stable Diffusion PNG or JPEG, click to choose, or paste
JPEG · PNG · WebP · AVIF · HEIC (Safari) — checked on your device
The most self-documenting AI images
Stable Diffusion is open source, so it is run in many front-ends — AUTOMATIC1111, Forge, ComfyUI, InvokeAI, Fooocus, SD.Next — and almost all of them save the generation settings into the PNG by default. Hobbyists rely on this: dropping a PNG back into the tool restores the exact recipe. For a detector, it means a typical Stable Diffusion PNG straight from someone’s computer contains a signed confession.
GPTTrace parses every PNG text chunk, inflates compressed ones, and checks whether the content looks like diffusion settings: a “Steps:” and “Sampler:” line, a seed, a CFG scale, a model hash, or a ComfyUI node graph with class_type entries. A match is reported as proof and the raw text is shown so you can read the prompt yourself.
When the settings are gone
Images that reach you through social media, image boards or a “save as JPEG” lose those chunks. Three signals remain.
Latent-grid dimensions
Stable Diffusion works in a compressed latent space eight times smaller than the image, and most workflows use sizes that are multiples of 64: 512×512 for SD 1.5, 1024×1024 and buckets like 832×1216 or 1216×832 for SDXL. A crop or an upscale hides this, but untouched outputs often keep it.
Decoder fingerprints
The VAE decoder that turns latents into pixels leaves a faint periodic structure. In the Spectrum view it can appear as a regular pattern of bright points — the “lattice” GPTTrace measures. Upscalers and JPEG compression weaken it.
The classifier
The neural model was trained on thousands of generators, including a large range of Stable Diffusion checkpoints and fine-tunes. It is the main signal for stripped images and the one most affected by heavy editing.
A note for people who make SD art
If you post Stable Diffusion work and want to avoid being accused of passing it off as photography or hand-made art, leaving the parameters chunk in your PNGs is the simplest form of disclosure. If you sell prints or entries to contests that require human-made work, check the rules: many competitions now ask for the generation metadata as part of the submission.
What Stable Diffusion and NovelAI leave in a file
These are the traces GPTTrace checks for. “Some files only” means the trace is often missing — a re-save, screenshot or social-media upload removes metadata — so its absence proves nothing.
| Generator | Where | What to look for | How reliable |
|---|---|---|---|
| Stable Diffusion | PNG tEXt “parameters” | Prompt, negative prompt, Steps, Sampler, CFG scale, Seed and Model hash (AUTOMATIC1111, Forge, SD.Next) | Documented by the vendor |
| Stable Diffusion | PNG tEXt “prompt” / “workflow” | The full node graph as JSON (ComfyUI) | Documented by the vendor |
| Stable Diffusion | PNG tEXt “invokeai_metadata” | Generation settings (InvokeAI) | Documented by the vendor |
| Stable Diffusion | Pixels | Sizes that are multiples of 64, e.g. 512×512, 768×768, 1024×1024, 832×1216 | Commonly seen |
| NovelAI | PNG tEXt | Software = NovelAI and a Comment JSON with the generation settings | Commonly seen |
| NovelAI | Alpha channel | A “stealth pnginfo” copy of the settings hidden in the alpha channel’s low bits | Commonly seen |
How accurate is it? Our measured numbers
We test GPTTrace on labelled image samples and publish the results, including where it does badly. It is tuned to keep false accusations rare, so it misses some AI content rather than flag real work.
image check: AUC 0.938 (cross-validated)
679 labelled samples (399 AI, 280 human), run 2026-10-08. At the “Likely AI” line it caught 64% of AI samples and wrongly flagged 5% of human ones.
| Source | Truth | Samples | Result at “Likely AI” |
|---|---|---|---|
| sdxl | AI | 40 | 100% caught |
| fullsize-photos | Human | 40 | 10% wrongly flagged |
| open-images-photos | Human | 40 | 5% wrongly flagged |
| lfw-faces | Human | 40 | 0% wrongly flagged |
| caltech-objects | Human | 40 | 3% wrongly flagged |
| coco-photos | Human | 40 | 0% wrongly flagged |
| ffhq-faces | Human | 40 | 0% wrongly flagged |
| celeba-faces | Human | 40 | 18% wrongly flagged |
Data sources and method: methodology & accuracy.