Skip to main content
Deploy’s AI-Generated Image & Video Detection API analyzes images and videos for signs of AI generation and manipulation. A single endpoint combines AI-generated media detection and deepfake detection, with an additional segmentation capability for images.

Detection Capabilities

AI-Generated Media Detection

Determine whether an image or video is AI-generated and identify its likely generative source.

Deepfake Detection

Identify AI-manipulated faces and return a confidence score indicating potential deepfake content.

AI Segmentation (Beta)

Locate regions of an image that show signs of AI manipulation and retrieve a segmentation mask.

AI-Generated Media Detection

The model determines whether an image or video was generated by AI. It is trained on millions of AI-generated and human-created images, including photographs, digital artwork, traditional art, and memes. The response provides:
  • Generation classification: Whether the content is AI-generated.
  • Source attribution: The likely generative model or platform.
  • Confidence scores: Scores for each classification.
  • C2PA metadata: Information about where the image came from, when available.

Classification Outputs

Supported Generative Models

The source classifier supports a wide range of image and video generators, including Midjourney, DALL-E, Firefly, Flux, Runway, Sora, Gemini, and others.
sora, pika, haiper, kling, luma, hedra, runway, hailuo, mochi, flux, hallo, hunyuan, recraft, leonardo, luminagpt, var, liveportrait, mcnet, pyramidflows, sadtalker, aniportrait, cogvideos, makeittalk, sdxlinpaint, stablediffusioninpaint, bingimagecreator, adobefirefly, lcm, dalle, pixart, glide, stablediffusion, imagen, amused, stablecascade, midjourney, deepfloyd, gan, stablediffusionxl, vqdiffusion, kandinsky, wuerstchen, titan, ideogram, sana, emu3, omnigen, flashvideo, transpixar, cosmos, janus, dmd2, switti, 4o, grok, wan, infinity, veo3, imagen4, krea, moonvalley, higgsfield, gemini, reve, heygen, sora2, seedream, seedance, grokimagine, gemini3, mai, lucid, sanavideo, flux2, qwen, zimage, viduq2, bria, blip3o, pixverse, ovis, longcat, cogview, bagel, ray3, gptimage1_5, hidream, steadydancer, ltx, personalive, dreamid, scail, meta, vibe, imagineart, seedance2, happyhorse.Additional classes:
  • other_image_generators: AI-generated content from an unlisted source
  • inconclusive: Source could not be confidently identified
  • inconclusive_video: No video source identified
  • none: No AI generator is associated with the content.
Source attribution is separate from AI-generation detection. Content may be identified as AI-generated even when the specific source cannot be determined.

C2PA Metadata

When available, the API also returns C2PA metadata containing information about the media’s origin, such as which AI model created it. Because metadata can be removed or falsified, we recommend interpreting it alongside the model’s detection results.

Deepfake Detection

The Deepfake Detection model identifies potential AI-based face manipulations in images and videos. It operates in two stages:
1

Face Detection

Identifies faces present in the input media.
2

Deepfake Classification

Evaluates each detected face for signs of AI manipulation and assigns a confidence score between 0 and 1.
The API returns the highest confidence score across all detected faces under the deepfake class. A score closer to 1 indicates stronger evidence of deepfake manipulation, while a score closer to 0 indicates weaker evidence.

AI Segmentation

Beta AI Segmentation identifies regions of an image that show signs of AI manipulation, including areas where objects have been added, replaced, or removed using AI. For an input image, the model can return:
  • Segmentation confidence: A score indicating evidence of AI manipulation.
  • Segmentation mask: A URL to a binary mask highlighting detected AI-manipulated regions.
Both the segmentation confidence score and mask can detect AI manipulation in partially edited and fully AI-generated images.
A high segmentation confidence score does not necessarily mean an image was partially edited. Fully AI-generated images can also receive high scores and produce segmentation masks.
We recommend the following thresholds as starting points for production workflows.
These thresholds are general recommendations. For optimal performance, calibrate them using representative data from your use case and your tolerance for false positives and false negatives.

Supported File Formats

  • GIF
  • JPG
  • PNG
  • WEBP

API Reference

This API is currently only available on our legacy API. Please refer to our AI-Generated Image & Video Detection API Reference for more details.