Templates are pre-made combinations of prompts and modifiers allowing you to quickly see a few different examples if you are new to diffusion models.
Controls the size of the generated images. Note that you can only change this parameter for Imagine. For other types of jobs, the size of the generated images is controlled by the size of the selected area in Photoshop.
Every output from Alpaca starts with a “prompt”, a textual description of the desired output. For example:
prompt: a house on hill
A prompt can be as simple or as complex as you like. Detailed prompts tend to produce better results.
Finding a good prompt that matches what we want is not always easy, especially for newcomers to the world of diffusion models. To make the process of finding good prompts easier, there are tools such as Lexica to see example of prompts and the kind of result you can expect by using them.
Note that using the same prompt as an online example, will not generate the same image. More on that in the seed section.
You can also decide to not use a prompt, by leaving this box empty. The model will then generate a result unguided.
Number of Images
We can specify the number of images we wish to generate from a particular prompt and set of modifiers. Each image will produce a different result.
Once your images has been generated, you will see the following appear in the Jobs panel. The arrows will allow you to toggle through the set of generated images. The check will allow you to select your current image and discard all others. The cross will discard all generated images.
Modifiers allow you to automatically modify a prompt by adding attributes or
qualities to it.
This is helpful to freeze part of a prompt once you found something you like so that you do not have to retype everything all the time.
A modifiers is composed of two parts: the text and the weight. The text simply describe the attribute you want to save, and can be as long as you want.
The weight describes how strong that attribute should be, the bigger the weight, the more your prompt will be skewed toward it.
Modifiers are stored in the format
<text>::<weight>, for example:
modifiers: manga::3 anime::1 drawing::1
We can generate variations on our basic house example from above by adding in some modifiers.
modifiers: monet::2 watercolor::1 sunrise::1
modifiers: simpsons::2 animation::1 forest::1
modifiers: art deco::1 modernist::1 photograph::1
The seed is an integer value that is the source of randomness in the model. Each image By varying the seed, we can generate different images from the same prompt and set of modifiers.
On the other hand, generating twice with the same set of prompt, modifiers & seed will generate the exact same result twice.
You can leave the prompt to 0 if you just want the model to use a random seed everytime.
prompt: a house on a hill modifiers: studio ghibli::2 seed: 1
prompt: a house on a hill modifiers: studio ghibli::2 seed: 24
Because the seed is responsible for the random noise that the model will then use as a starting point to generate images, using the same seed, even when varying prompts and modifiers will result in images that have the same qualities, such as the general color palette, placement of objects etc. You can use this fact to your advantage to increase the coherency of different generations.
This controls the number of diffusion steps that carry us from the original source noise to the final image. By increasing the number of steps, we generally improve the quality of the result and it’s correlation with the prompt and modifiers.
The prompt strength controls the influence of the prompt — higher values of this parameter force the model to be more attentive to the prompt. We can consider this to be a “rigid” vs “relaxed” interpretation of the prompt.
prompt: a house on a hill modifiers: disney::2 animation::1 seed: 1 prompt strength: 2
prompt: a house on a hill modifiers: disney::2 animation::1 seed: 1 prompt strength: 11
The transfer strength controls how much we want to modify the input image. High transfer strength will modify the input image strongly, resulting in an image that may not have much relation with the input anymore. While a very low strength will not leave enough time for the model to make any noticeable changes to the image.
A balanced strength, will keep the main features of the input image alive while modifying it enough so that it becomes closer to your prompt