Preamble

I have over 1000 and less then 2000 trainings now. I am far from an expert in any subject regarding AI.

I have a fair aptitude for pattern recognition (Or maybe I just stare at the screen to long) and something has been bothering me about AI for a long time. Its a pattern I see, but not a natural one like Mandelbrot. I will try to site sources but largely this is my conjecture.

512X512

A 512x512 grayscale image is made up of 262,144 pixels.

So if your taking black and white photos of any given subject the possible number of combinations on a lossless (Uncompressed PNG, BMP, Targa) image is as follows:

262,144 with an Exponent of 256 which is a number larger then a google.

The calculation for a color image would be the number of pixels in a given resolution to the exponent of 16 million.

So we have a number of images that is technically quantifiable but beyond comprehension, practically an infinite possibility in training a model right?

PRNG

Pseudo Random Noise Generation, this is a term many in AI might not be familiar with but it effects every image you generate.

A seed is just a table to call a "random" number that makes the noise your image starts with. It is repeatable because it is not random.

In fact the number of seeds available is very limited.

• 2 billion for 32-bit seeds

• 65,536 for 16-bit

• 255 or less for FP8

Why PRNG Matters

Your video card uses a technique called Xorshift to generate noise (With CUPY or NUMPY calling on cuRAND)

With an already limited number of "Random" numbers, it uses a algebraic function that further reduces its quality.

The ability to generate images from noise and the reverse of that process is a central function of generative models called diffusion

While using a PRNG with a limited number of combinations can lead to a more controlled exploration of the latent space, it can also introduce biases. If the randomness is not sufficiently diverse, the model may struggle to explore different configurations effectively.

Is RDRAND the answer

On an Intel I7 it would take just under 1 second to generate a truly random number on a 512x512 image.

It would take approximately 86.51 milliseconds to perform 786,432 calculations at 110 nanoseconds per calculation.

Seeds and known noise values are an integral part of diffusion model training, and I am not sure if the 2 billion limitation and XOR shift is causing the patterns I am seeing.

Other possibility I have considered regarding these same familiar faces and patterns in AI models.

• Fourier Transform and the losses caused by JPG or up scaling

• XFORMERS or Flash Attention

• Hardware "shorcuts" see Dept of Energy testing AMD cards"

Impact Statement

Its possible that in training a 1T model, teaching it from a "random" 2 billion seeded images with a 3 pixel shift in noise is sufficient complexity.

We have been using mathematical compression techniques in things like JPG for 30+ years. The detrimental effect they have on complexity while not readly visable to the human eye could be magnified over trillions of iterations.

A collection of images that have not been touched by Fourier or Lanczos would be priceless.