Image Seed Control
FooocusPlus and other generative AI interfaces typically generate images using a reverse diffusion process. This process starts with an image composed of random noise that is intelligently refined into a coherent image. The initial noisy image is generated using a random noise generator and that noise generator must be seeded with a numerical value.
Using the same prompt (including styles) and the same image processing parameters, a particular seed value will ultimately produce an identical image. Basic image diversity - the ability to produce a variety of images - is introduced by starting with a different seed value, even while the prompt and all the other processing parameters remain the same.
To maximize diversity, it is important that the seed value is chosen in a truly random way. While personal computers typically use a pseudo-random number generator that does not achieve the highest possible degree of randomness, the randomness they supply is good enough for the generative process.
By default, FooocusPlus starts with Random Seed selected, as shown above. In this mode a new random seed is chosen at the beginning of a generative batch. For each image created within the batch the seed is incremented by one. Adding one to the initial seed value should create a new and completely random image.
In this example, an Image Grid was created using four random images. The Image Log shows that the first image used a seed of 7509075761779731798. After each image generation the seed was incremented so that the last image had a seed of 7509075761779731801.
What if we wanted to change the prompt and see how this batch turned out?
When the Random Seed checkbox is deselected, the Specific Seed box appears. The seed number shown is the first number used in the last batch of images. In this case the number is 7509075761779731798 as revealed above. This number could be changed by typing a new 19 digit number into the box but in this case we will leave it as is.
The first batch used this prompt:
a __canadian__ man and a __canadian__ woman smile in a New Year's greeting on the street of a __canadian_city__
A simple prompt using three wildcards and the Default preset.
But using the same sequence of seeds, this example shows the results of changing the canadian_city wildcard to the nationality wildcard plus the word "city":
a __canadian__ man and a __canadian__ woman smile in a New Year's greeting on the street of a __nationality__ city
Our four sets of subjects have been transported elsewhere to celebrate New Year!
Using a Specific Seed is the key to maintaining Character Consistency. With Freeze Seed enabled the Specific Seed will not increment for each new image in a batch. By using wildcards or a prompt array, the same subject or subjects can be shown in different situations, with different clothing, or doing different things.
For this set we are using the description and seed for the initial pair with a prompt array containing four options:
a Multi-Ethnic Canadian man and a English Canadian woman [[run through the forest | walk across the desert | embrace on a mountain top | dance on the beach]]
In theory, a frozen seed should enable a batch of images to display the same subjects in different scenarios. However, we are using an SDXL model and the inconsistent results in the example demonstrate that character consistency typically requires the use of a more advanced model such as a one in the Flux1 Dev series, as illustrated in Character Consistency & Storyboards.
With the Extra Variation option, the seed for each image in a batch is not incremented by one. Instead a new nonconsecutive random number is chosen for each image. The big disadvantage to this is that it is impossible to automatically repeat the sequence if so desired. As to whether using a nonconsecutive number actually causes more randomness, that is a subject for debate. While I understand the arguments that this does not cause extra randomness - even though we are dealing with pseudo-random numbers rather than truly random ones - there are times when I have been pleased to use this option when I become impatient with seeing the same patterns repeat over and over again.
Like Frozen Seed, Extra Variation can be used in either the Random Seed or Specific Seed modes. In the final example, Extra Variation is used with the original prompt and the same specific seed. The first image is therefore identical to the original example, but the remaining three images are quite different.
