AI image detector by generator
SDXL image detector: prompt, style and model from the PNG
SDXL images are identified from the records their interface wrote into the PNG, and isGenAI reads the saved prompt, the recorded checkpoint name and SDXL-specific settings such as the refiner and style prompts when they are consistent.
Illustration · PNG file
- Signature + IHDR
- iTXt XML:com.adobe.xmprecord
- tEXt parameters
- tEXt prompt / workflow
- tEXt invokeai_metadata
- eXIf
- caBX C2PA manifest
- IDAT pixels
- IEND
AI source type declared
The file labels itself as AI-made.
IPTC DigitalSourceType such as trainedAlgorithmicMedia is read from XMP and shown with its exact value.
Check an image →What an SDXL workflow leaves behind
SDXL is Stability AI’s larger open-weight model. The base-plus-refiner pipeline and the separate style prompt mean the saved record is richer and more varied than a classic Stable Diffusion 1.5 file. AUTOMATIC1111 writes the whole thing into the parameters text chunk, including the model hash and refiner fields. ComfyUI saves the prompt graph with both samplers. InvokeAI’s invokeai_metadata record carries sdxl_txt2img as its generation mode and may hold positive and negative style prompts.
The model itself writes nothing. A hosted SDXL service may strip the record or add its own; check the actual download.
How isGenAI reads SDXL records
A single-sampler graph or a complete parameters chunk yields a saved generation prompt with the model name and settings. A two-stage ComfyUI graph with base and refiner samplers is reported as a saved workflow, because two samplers mean two candidate prompts and the checker refuses to guess which one applies. InvokeAI records with a style prompt that differs from the main prompt are also reported as a workflow rather than a single instruction.
This caution is deliberate: a report that names one prompt must be able to point at exactly one consistent record in the file.
Formats and reposts
PNG keeps the records; a JPEG or WebP saved by AUTOMATIC1111 keeps them in EXIF UserComment. Anything re-encoded by a browser, an editor or a social platform loses them. Use the image metadata compare tool to see which fields disappeared between the original and a copy.
A missing record is not evidence of a human author. isGenAI prints no evidence found in that case.
How to check an image from this generator
- Get the original download, not a screenshot or a copy saved from a chat or social feed.
- Upload it to the free AI image detector. Files up to 20 MiB in PNG, JPEG, WebP, AVIF or TIFF are read on the server and are private by default.
- Open “How we checked” and read each finding with its source field: signed record, declared source type, saved prompt, AI statement or matching copy.
- Treat a clean result as “no evidence found”, never as proof that a person made the picture.
Frequently asked questions
Does SDXL add metadata to images?
The model does not. The interface running it does: A1111 writes a parameters chunk with refiner fields, ComfyUI writes prompt and workflow chunks, InvokeAI writes an invokeai_metadata record with sdxl generation modes.
Why does my SDXL image show a workflow but no prompt?
Base-plus-refiner graphs contain two samplers, and InvokeAI records may carry a separate style prompt. isGenAI reports a saved workflow instead of guessing which prompt produced the final pixels.
Is an SDXL image detectable after a screenshot?
No. Screenshots and social uploads remove PNG text chunks and EXIF, and isGenAI does not infer origin from pixels.
Can isGenAI tell SDXL from Stable Diffusion 1.5?
Only from the saved record: the recorded checkpoint name, size and generation mode. Without a record the checker makes no claim about the model.
Sources
Vendor behaviour changes. These pages describe what the vendor documents and what isGenAI reads from a file; they do not test every release.