How AI Image Provenance Works: From Prompts to Manifests

## Where AI Generators Leave Metadata Footprints When a neural network generates an image, the software pipeline frequently bundles metadata into the resulting file. Depending on the software used, these footprints take several different forms. --- ## 1. Open-Source AI Generators (Stable Diffusion, ComfyUI, AUTOMATIC1111) Open-source generation tools prioritize workflow reproducibility. Consequently, they embed comprehensive generation state into the file headers: - **PNG Text Chunks (`tEXt`, `iTXt`):** - `parameters`: Full positive and negative text prompts, CFG scale, sampler name, step count, and random seed. - `workflow`: The complete ComfyUI node graph serialized as JSON. - `sd_checkpoint`, `lora`: Hashes and filenames of models used during inference. --- ## 2. Commercial Cloud Generators (Midjourney, DALL·E, Adobe Firefly) Commercial platforms use different methods: - **Adobe Firefly:** Embeds C2PA Content Credentials recording the `com.adobe.generative-ai` assertion and signing certificate. - **DALL·E / ChatGPT:** Embeds C2PA manifests into JPEG and WebP outputs declaring provenance via OpenAI credentials. - **Midjourney:** Embeds metadata in PNG text chunks, XMP creator tools, and sometimes IPTC credit fields. --- ## 3. Metadata vs Watermarking vs Visual Artifacts It is vital to distinguish between three layers of AI identification: ``` +-------------------------------------------------------------+ | Layer 1: File Metadata (EXIF, XMP, PNG Chunks, C2PA) | | -> Easily inspected & stripped via client-side tools | +-------------------------------------------------------------+ | Layer 2: Invisible Steganographic Watermarks (SynthID, etc.)| | -> Embedded imperceptibly within frequency domain pixels | +-------------------------------------------------------------+ | Layer 3: Visual / Statistical Classifiers | | -> Neural networks trained to spot generative pixel textures| +-------------------------------------------------------------+ ``` MetaClean operates exclusively on **Layer 1 (File Metadata & Provenance Headers)**. Cleaning metadata removes file header traces, but does not alter invisible pixel-domain watermarks.