AI-Generated Images: Navigating Innovation and Misinformation

In the intersection of science and art, AI’s ability to visualize our imagination has become a significant breakthrough. This blog post examines the straightforward process of generating synthetic images using AI, specifically focusing on systems like DALL-E.  Because of the ease with which these images are generated, we emphasize the need for a collaborative effort within the community to establish compliance standards that would prevent the spread of misinformation.

The Prompt:

The first step in this process is writing a detailed prompt, providing the AI with the necessary information to produce the desired image. For example:

Prompt 1: “Photo of a multi-stained immunofluorescence sample, with DAPI stained blue fluorescent nuclei complemented by additional stains: red for actin fibers and green for a specific protein of interest. The diverse colors illuminate distinct cellular components against a dark background, providing a complex and informative view of the interplay between different cell structures.”

Using this prompt, the AI, which has been trained on a vast dataset of images and text captions, generates an image that closely aligns with our specified requirements.

The Seed:

The concept of randomness in AI image generation is rooted in deterministic processes. Computers utilize a pseudo-random number generator (PRNG) for tasks that require randomness, such as AI image generation. The PRNG is essential for the diversity of the output, creating a unique image each time. However, this process is not truly random; it is deterministic, meaning that the same initial conditions will produce the same output every time.

Seed values are paramount in this deterministic system. A specific numerical seed initializes the PRNG, dictating the pseudo-random sequence that influences the minute aspects of the resulting image, such as strokes, patterns, and colors. Using the same seed value allows us to reproduce identical images, which is vital for revisiting and sharing specific digital creations.

Synthetic images generated with Prompt 1. Initial seeding number is 524,459,761 and increased by 1; images correspond from top left to bottom right.

Figure 1:  Synthetic images generated with Prompt 1. Initial seeding number is 524,459,761 and increased by 1; images correspond from top left to bottom right.

There are technical limits to the seed values. For instance, in a 32-bit system such as this model relies, seed numbers can range between +/- 2,147,483,648 (or 232 integers).  Infinite images of the same prompt cannot be produced, as exceeding this value is not possible as it would require a larger data type to store the number. It’s crucial to understand that reproducibility with seed values assumes an unchanged model. If the underlying AI model is updated or retrained, the same seed value may yield different outcomes due to the updated knowledge and parameters within the AI.

Synthetic images generated with Prompt 1. Images correspond with the minimum and maximum seeding numbers.

Figure 2:  Synthetic images generated with Prompt 1. Images correspond with the minimum and maximum seeding numbers.

The Consequences …

AI image generation, with its balance between algorithmic precision and controlled randomness, is a powerful and sophisticated form of artistry and should be celebrated. Unfortunately, access to this technology in its current form has the potential to propagate misinformation—intentionally or not.

Urgency within the community must be cultivated to ensure images are readily detectable as synthetic, or ‘fake’, to avert its misuse in scientific publications and public distribution. Collective efforts must be made to enhance watermarking techniques, metadata embedding, and steganography methods (the ability to hide images/data within an image) in order to create vigorous standers for current and new AI-generative models.

It is the responsibility of researchers, publishers, and AI developers to collaborate towards compliance standards when this technology is still in its early stages, safeguarding against these tools as their sophistication escalates.

Authors Note:  This article was written on November 3rd 2023, when the seeding numbers were still accessible to users in order to regenerate images under the same conditions. After the most recent update to ChatGBT, accessibility to this feature has dropped; PRNG numbers can still be obtained but can no longer be used as an input. 

If you’re interested in working with Visikol, please reach out to a member of our team today!


Nicole Callen
Senior Image Analyst

2023-11-28T12:24:35-05:00Tags: , , , |

Share This Page, Choose Your Platform!


This website uses cookies to enhance the user experience. Ok