Page 99 - Libro 2

ARTICLES ARTICLES

6. A closeup examination reveals unrealistic discrepancies and strange artifacts.

Given the accuracy issues when performing a
Subjective Image Assessment, and the possibility of
images such as compositive photos having similar
inconsistencies, I have found it often necessary
to include an Objective Technical Analysis using
impartial quantitative metrics. These metrics
identify patterns, attributes, and anomalies. These
metrics also examine distinctive features, colour
values, resolution, and noise within images.
Oddities Contained In addition, other tools used in this analysis
included image metadata analysis and reverse
image searches. All of these processes use
standardized techniques for evaluating images.
In the tests conducted by our committee, two
Movie Marque
categories of AI generative images were analyzed.
These two categories were first, images generative using AI text-to-images, and second Images
that were enhanced using Adobe Photoshop 2024’s generative AI features (generative fill/remove
elements and expand boundaries of an image).
7. Motor vehicles may seem unrealistic, contain license plates and vehicle identifying lettering
that are nonsensical. Note that AI DeNoise, AI Sharpening and AI enlarging fall under the category of Machine
Learning, a subset of Artificial Intelligence. They employ algorithms and statistical models to enable
8. The image contains text that does not accurately reflect the visual content. computers to learn and execute specific tasks. However, their dataset is considerably smaller than
what is required for AI-Generative technology. Hence, AI DeNoise, AI Sharpening and AI enlarging
do not receive the label of an AI-generated image in the metadata file.
To perform an objective technical analysis 100 AI generated images that reflected a wide range
of genres were created in each of the following nine AI generators:
• Adobe Firefly 2 text-to-image
• Bing AI
• Dream Studio
• Leonardo AI
Articles • Meta AI Articles

Midjourney
•
NightCafe
•
• SeaArt
• Stable Diffusion

Unrealistic and nonsensical lettering All of the 900 AI generated images that were created from the above listed generators as well
as 100 images that were captured in-camera and submitted to a 2020-2021 national exhibition in
Canada were tested using two AI classifiers. The classifiers that were used to test all images were AI
or Not and Hive AI Detection. Both AI or Not and the Hive AI Detection websites function as machine
Some of these visual inconsistencies can be seen in the image of the movie marquee generated learning services, analyzing the surface content of images to determine whether each image
in Dream Studio AI. Many of these imperfections will not always be obvious as they can be originated from a photograph or was generated by an AI generator.
corrected by refining the generative AI prompt, employing negative prompts, and subsequently
regenerating the prompt. Currently, methods to mask detection of AI generation include using The 1,000 image database, consisting of the 900 AI rendered images from the nine different AI
AI techniques like in-painting to replace generative text, specifying the exact text in double- generative text-to-image creations and the 100 camera-captured images, were tested by AI or Not
quotations within the AI prompt, and using a post-processing application to insert the required and Hive AI Detection. With the exception of Adobe Firefly text-to-image creations, both AI or Not
text into the generated image. and Hive AI Detection exhibited a high degree of accuracy in correctly identifying photographs from
generative AI text-to-image creations.

98 99