Which AI made this image?
Upload an image to identify the AI generator behind it. RAIW checks signed provenance metadata and visible marks first, and analyzes image pixels across ten AI generator families when metadata has been stripped.
How RAIW identifies the AI generator
RAIW checks your file in two steps. First, it inspects embedded provenance records and supported visible corner labels: signed Content Credentials (C2PA), IPTC AI-source tags, generator EXIF, XMP and PNG parameters, and China AIGC (TC260) disclosures. When a record is intact, RAIW names the generator directly from the file.
When metadata is missing after a download, re-save or social upload, RAIW analyzes the image pixels on files up to 9 MP and compares them against calibrated generator pipelines. When the pixels do not match any supported generator closely enough, it abstains and reports that it could not tell.
Supported AI generators and model families
- OpenAI - ChatGPT and API exports across GPT Image 1.5, GPT Image 1, DALL·E 3 (including Microsoft Copilot), and DALL·E 2, plus OpenAI C2PA manifests.
- Google - Nano Banana 2 (Gemini 3.1 Flash Image), Nano Banana Pro (Gemini 3 Pro Image), Nano Banana (Gemini 2.5 Flash Image), and Imagen 3, plus Google C2PA provenance and the visible Gemini sparkle.
- ByteDance - Doubao, Jimeng (Dreamina), and Seedream from pixel pipelines, TC260 labels, C2PA manifests, and visible corner marks.
- FLUX (Black Forest Labs) - FLUX.1, FLUX.1 [dev], FLUX.1 [schnell], FLUX.2 [dev], and FLUX.2 [pro], plus BFL C2PA credentials.
- Stable Diffusion - Stable Diffusion XL (SDXL), Stable Diffusion 3, 2, 1.5, and 1.4, plus Automatic1111 and ComfyUI PNG generation parameters and Stability AI C2PA manifests.
- Midjourney - Midjourney v6.1 and v5.2 pixel pipelines, plus Midjourney XMP and PNG metadata when preserved.
- Ideogram, Adobe Firefly, Meta AI, and xAI Grok - Ideogram 2.0 and 1.0, Adobe Firefly Image 3 and Firefly C2PA manifests, Meta AI Imagine, and xAI Grok Imagine and Grok 2 (Aurora).
- Provenance-only and visible-mark providers - when signed metadata or a supported visible mark is present, RAIW also attributes Qwen, Kling, Wan, Tencent Yuanbao, Dola AI, WorkBuddy, Samsung Galaxy AI, Apple Photos Clean Up, Apple Image Playground, Canva, Runway, Recraft, Luma AI, Leonardo AI, Bria AI, fal.ai, Amazon Bedrock (Nova), Higgsfield, Suno, Producer.ai, Sync Labs, PixelBin.io, Topaz Labs, CyberLink PhotoDirector, and SPRING.
What pixel attribution can and cannot prove
Pixel analysis identifies a likely AI generation and export pipeline, not an invisible watermark. It does not locally decode SynthID, prove that a hidden watermark is present, or certify that an image was taken by a camera or made without AI.
Camera photos, UI screenshots, vector graphics, diagrams, heavy edits, and generators outside the ten calibrated families normally return a neutral could-not-tell result. To check for a vendor's own invisible watermark, use the Google Gemini app or openai.com/verify.
How to check an image
- Upload a PNG, JPG, WebP, AVIF, HEIC, or HEIF image (up to 120 MB). No account needed.
- Read the plain-language verdict: what was found and the likely source.
- If a watermark or label is found, choose Continue to removal to clean the same image.
Related tools
Frequently asked questions
How can I tell which AI made an image if metadata was stripped?
Can this tool prove that an image was taken by a camera?
Does identifying Google or OpenAI from pixels mean SynthID was detected?
Why did the checker say it could not tell which AI made my image?
Is checking free? Do you store my image?
What image formats can I check?
Checking is free and needs no account.