Stable Diffusion Detector: How to Identify Diffusion Model Images
Stable Diffusion and its derivatives (SDXL, SD 3, ComfyUI, Automatic1111, Forge, and thousands of fine-tuned community models) leave consistent forensic traces: specific output dimensions, no camera EXIF, and characteristic pixel smoothness in flat regions.
Stable Diffusion is an open-source diffusion model that runs locally on consumer hardware, making it the most widely deployed AI image generator in terms of total installations. Unlike closed-platform generators, Stable Diffusion images can be generated with no platform logging, no usage policies, and no watermarking. This makes Stable Diffusion the tool most frequently used to generate non-consensual deepfakes and explicit synthetic imagery. However, the open-source nature has also created documented, consistent technical signatures.
Stable Diffusion metadata signatures
Images generated by Stable Diffusion frontends (Automatic1111, ComfyUI, SD.Next, Forge) often embed generation parameters in the PNG metadata, including prompt text, model name, sampling method, and CFG scale. These parameters are stored in the 'Comment' or 'parameters' PNG text chunk. A PNG file with prompt text in its metadata is definitively AI-generated. When this data is present, it explicitly names the model and often includes the LoRA or fine-tune being used.
Stable Diffusion output dimensions
Stable Diffusion 1.x uses a 512x512 default resolution (or 512x768, 768x512 for portrait/landscape). SDXL uses 1024x1024 as default, with 1152x896, 896x1152, 1344x768, 768x1344, 1216x832, and 832x1216 as standard alternatives. SD 3 supports up to 2 megapixel outputs but still defaults to 1024x1024. All outputs are multiples of 64. Img2img workflows can produce images at non-standard sizes but still show characteristic pixel smoothness.
AUTOMATIC1111 and ComfyUI detection
AUTOMATIC1111 and ComfyUI (the two most widely used Stable Diffusion frontends) embed full generation metadata by default. AUTOMATIC1111 encodes this in the PNG 'parameters' text field in a specific format: prompt, negative prompt, and generation settings on separate lines. ComfyUI embeds the entire workflow JSON in the PNG metadata. This metadata is the single most reliable detection signal for locally-generated SD images, but it is removed when the image is screenshot or saved as JPEG.
Detecting Stable Diffusion images without metadata
When metadata has been stripped, either deliberately or through format conversion, detection relies on dimension patterns and sensor noise analysis. Sensor noise analysis is particularly effective against Stable Diffusion outputs because the model generates very smooth flat regions with near-zero pixel noise. Real cameras produce 2 to 5 luminance units of noise even in uniform areas; SD outputs typically show less than 0.5 units in comparable regions.
Check any image with the free detector
Upload a photo to analyze its metadata, sensor noise, and pixel patterns. No account required. Your image never leaves your device.
Use free detectorFrequently asked questions
Can Stable Diffusion generate deepfakes?
Yes. With face-specific LoRA models and DreamBooth fine-tuning, Stable Diffusion can generate realistic images of specific individuals. This is the primary technical basis for most consumer-level deepfake generation.
Does the ScanErase detector work on locally-generated Stable Diffusion images?
Yes. The detector checks for SD standard dimensions, sensor noise absence, and PNG metadata parameters. An unmodified AUTOMATIC1111 or ComfyUI output will typically show high AI probability scores.
Are Stable Diffusion deepfakes covered by federal law?
Yes. The TAKE IT DOWN Act applies to all AI-generated intimate imagery regardless of the generation tool, including open-source models running locally.
Found a deepfake? Find every copy.
Detection is step one. ScanErase's biometric scan finds every copy across 200+ platforms, including social media, for $15, covering every platform simultaneously.
Start your scan