Synthetic Image Evaluator

​

Evaluation Using Interface

• Required Inputs:
  • input_image: URL or file path to the image to be evaluated.

• Result: Returns ‘Passed’ if the image is detected as AI-generated, ‘Failed’ if it appears to be a real photograph or human-created image.
• Reason: A detailed explanation of why the image was classified as AI-generated or not.

​

Evaluation Using Python SDK

Click here to learn how to setup evaluation using the Python SDK.

• Required Inputs:
  • input_image: string - URL or file path to the image to be evaluated.

• Result: Returns a list containing ‘Passed’ if the image is detected as AI-generated, or ‘Failed’ if it appears to be a real photograph or human-created image.
• Reason: Provides a detailed explanation of the evaluation.

result = evaluator.evaluate(
    eval_templates="synthetic_image_evaluator",
    inputs={
        "input_image": "https://www.esparklearning.com/app/uploads/2024/04/Albert-Einstein-generated-by-AI-1024x683.webp"
    },
    model_name="turing_flash"
)

print(result.eval_results[0].output)
print(result.eval_results[0].reason)

['Passed']
The evaluation is 'Passed' because the image displays clear indicators of AI generation.

*   The image shows **telltale signs of AI generation** in its rendering of Einstein's face, particularly around the eyes, which have an unnatural symmetry and lack the subtle imperfections found in photographs.
*   There are **inconsistencies in texture detail** between different parts of the image, with some areas showing hyper-detailed skin pores while others appear smoothed or blurred in an unnatural way.
*   The **background elements** have the characteristic softness and dreamlike quality common in diffusion-based AI image generation.
*   The **lighting effects** show physically implausible interactions with the subject's features, creating an uncanny appearance.

These characteristics strongly suggest the image was created using AI image generation technology rather than being a genuine photograph.

​

What to do If you get Undesired Results

• For actual photographs mistakenly identified as synthetic:
  • Ensure the image has not been heavily processed or filtered
  • Check that the image doesn’t have unusual artifacts from compression or editing
  • Consider providing higher resolution versions if available
• For synthetic images not being detected:
  • Be aware that newer AI generation models are becoming increasingly photorealistic
  • Some AI-generated images that were post-processed or combined with real photographs may be harder to detect
  • The evaluation works best with full images rather than small crops or heavily modified versions

​

Comparing Synthetic Image Evaluator with Similar Evals

• Caption Hallucination: While Synthetic Image Evaluator determines if an image was artificially created, Caption Hallucination evaluates whether descriptions of images contain fabricated elements not visible in the image.
• Toxicity: Synthetic Image Evaluator focuses on the creation method of images, whereas Toxicity evaluates whether content contains harmful elements.