Nebula AI Detector
Multi-model · Images not retained
Deciding whether an image was made by AI usually comes down to two angles: what the file metadata says, and what the pixels themselves look like. Upload an image to Nebula AI Detector and you get both at once, along with a 0-100 AI-probability score.
Many AI art tools write generation details into the file on export. Common traces include AIGC content labels (the TC260 standard), generator signatures such as Midjourney, Stable Diffusion, DALL·E, ComfyUI or NovelAI, and parameters like sampler, step count and prompt fragments. Conversely, camera make and model information (Canon, Sony, Apple iPhone, Huawei and so on) together with shooting parameters points towards a real photograph.
AI-generated images often differ from real photographs in fine detail, texture and noise distribution. Nebula runs a deep-learning model over the pixels and outputs an AI-generation probability, so it works even when the metadata has been stripped.
Metadata is the most direct evidence, but social platforms and screenshot tools strip it easily. Pixel analysis survives metadata loss, but depends on model judgement. Nebula presents the two independently so you can cross-check: a metadata hit on a known generator signature is conclusive, and when metadata is missing the pixel analysis still gives you an answer.
Privacy: images are used only for this detection and deleted immediately afterwards — never stored or cached. Results stay in your own browser.
There are two practical angles. First, inspect the file metadata (EXIF/XMP) for AIGC markers or signatures left by generators such as Midjourney or Stable Diffusion. Second, analyse the pixels themselves — AI images often differ from real photographs in fine detail and noise distribution. Uploading to Nebula AI Detector gives you both analyses at once, plus a 0-100 AI-probability score.
Some generators embed information when exporting, such as AIGC content labels, the tool name, and sampling parameters. However, this metadata is easily stripped by social platforms or screenshot tools. So a metadata hit is strong evidence, but when it is missing you need pixel-level analysis to reach a conclusion.
Signatures left by common generators are all recognised, including Midjourney, Stable Diffusion, NovelAI, DALL·E, ComfyUI and Flux. Beyond reading signatures, Nebula also runs a model over the pixels themselves, so detection does not depend on whether the metadata survived.
Yes, the image is uploaded for detection. It is used only for that detection, passes through an image host and auto-expires within 5 minutes. Our server never stores the original. Results are kept only in your own browser.
PNG, JPG/JPEG, WEBP, BMP and TIF/TIFF, up to 30MB per image.