This is a set of wildcards that enable random variance for the weights of LoRAs and Textual Inversion embeddings.

Use for LoRAs may look like <lora:my-lora:__proc/[modifier]__> where [modifier] can include the following:

• full: For when you want a lot of variance for common models. The range is 0.10-0.95 with an average of 0.60.

• high: High range values typical for most character/clothing/subject models. R:0.55-0.95, A:0.75

• mid or med: Mid range values. R:0.3-0.7, A:0.50

• low: Low range values for accents and themes. R:0.10-0.45, A:0.27

For a TI embedding, an example may look like (my-textual-inversion:__proc/[modifier]__) where [modifier] might be:

• buff: Enhances an embedding. R1.05-1.3, A:1.17

• dep: Depresses an embedding. R:0.80-0.95, A:0.88

There's also a series of values for weights that have a +/- 0.15 variance for finer precision. These can be accessed with the wildcard: __proc/p[value]__ where [value] can be a quarter {25, 50, 75} or a quintile {20, 40, 60, 80}. These precision weights are more flexible in their formatting because they do not include a digit before the decimal point, so you can adjust their magnitude by adding this. For example having 1__proc/p25__ in your prompt would return values ranging from 1.1 to 1.4. To make weights negative, add a minus sign in front of the wildcard.

Precision values that are a multiple of 100 are not as flexible because they must have a digit before the decimal point. __proc/p000__ can also return negative values, so adding a minus sign would create a double negative.

LoRA Block Weights:

You can also use wildcards to randomly assign lora block weights. The syntax is a little different depending on whether you're applying it to a SD LoRA, XL LoRA or LyCORIS.

SD: <lora:[my-sd-lora]:1,1:__proc/lbw/[modifier]__>

XL: <lora:[my-xl-lora]:1,1:__proc/lbwXL/[modifier]__>

Lycoris: <lora:[my-lycoris]:1:1:__proc/lbwLyc/[modifier]__>

The following modifiers apply to all three forms. Only UNet blocks are affected, the BASE block is always set to 1 using these methods.

• rand: Sets each individual block to 1 or 0.

• randIN: Same as rand except it applies only to IN blocks.

• randOUT: Same as rand except it applies only to OUT blocks.

• midd: Sets the MID and a random number of adjacent IN and OUT blocks to 1.

• wrap: Inverse of midd, sets MID and a random number of adjacent blocks to 0.

Other:

Finally, you can generate some whole numbers with the following dice. This could be useful in some edge cases with a little formatting.

• __proc/d10__ will return digits from 0 to 9.

• __proc/d100__ will return numbers from 00 to 99.

Cover art generated using SirDigsbey's model of Mary Wiseman.