Incredible artificial intelligence Easy Diffusion 3.0. The Middle and High School Guide. Alexander Chesalov. Читать онлайн. Newlib. NEWLIB.NET

Автор: Alexander Chesalov
Издательство: Издательские решения
Серия:
Жанр произведения:
Год издания: 0
isbn: 9785006270824
Скачать книгу
you are creating:

      For our first image, we will only add a description and select a model to base this image on. We will not change any parameters now.

      Please also note that above the Toolbar there are three additional buttons that allow us to add a description from a text or structured file, as well as use additional image modifiers:

      – «Load from a file». This button allows you to load text describing a previously created image from a text file or a JSON file. It is very useful when you are working with a large number of images. Previously, in the Settings window, we enabled the «Auto-Save Images» option and defined the «Metadata format» as the text «txt». This way our system saves all your images and descriptions in text format automatically.

      • «Embedding». Configured through the «Models tools» tab of the application’s working window.

      • «Image Modifiers» / «Image modifiers». A very useful button, especially at first, when you have not yet learned how to compose complex descriptions of new images. In the modifiers you can choose several different filters for your images:

      Below the buttons there are two windows:

      – «Enter Prompt». This is the main window, which is intended to contain a text description (we will call it briefly a «query») of the image that we want to create. We will work with him throughout our reading of the book.

      • «Negative Prompt». This is an additional window in which everything that should not be displayed in the created picture is described in short phrases.

      Below the request windows there are two buttons, united by one heading «Initial Image (img2img)»:

      – «Browse». This button allows us to add our image, which the Easy Diffusion system uses to create a new image similar in design style and presentation form from it and the request we wrote.

      – «Draw». The button opens the image editing window. In this window we can edit the image we added as a source. For example, clear the background of an image.

      Below is the «Make image» button, which is very important for us, with which we will start the process of creating our new unique images.

      Under the «Make image» button there is a large list of parameters for the image we want to create.

      Let’s take a closer look at all the «Image Settings» functions:

      – «Seed». This is the starting number of the random number generator that is used to create a unique image. By default, the program itself creates a unique number. If we know the Seed and other things being equal, we can create the same picture again. We do not change this option and leave it as it is.

      – «Number of Images». This option determines the number of images that are created after clicking the «Make image» button. We leave the default value (equal to «1»). If you change the number, it will affect the overall performance of the system. We do not change this option and leave it as it is.

      – «Model». This window displays a list of models that we downloaded for you earlier. Models are used to create unique images. We will look at their work later in this and subsequent chapters.

      – «Clip Skip». This option allows the system to more accurately analyze and interpret our request. When the option is disabled, then the new image created matches our request as closely as possible. We do not change this option and leave it disabled.

      – «ControlNet Image». This is a new and very interesting option of the Easy Diffusion Artificial Intelligence system version 3.0. With ControlNet, you and I can not only create a unique image, but we can also make it look like a celebrity. ControlNet will be covered in more detail in a separate chapter.

      – «Custom VAE». This option allows us to use a special autoencoder, which helps speed up the creation process and improve image quality. The Easy Diffusion manual suggests downloading and using the file vae-ft-mse-840000-ema-pruned.ckpt. I must say that it worked great with Easy Diffusion 2.5, but does not work correctly with our version of Easy Diffusion version 3.0. We do not change this option and leave it disabled.

      – «Sampler». These are auxiliary Text-to-image AI software for improving the created image. The interesting thing is that some samplers are suitable for certain types of images, while other samplers are suitable for other types of images. You will have to determine which one is better during your work.

      – «Image Size». This option allows you to set the size of the resulting image. I highly recommend that you use the 512x512 size. After you select the one you like from the many images you have created, you can use the «Upscale» / «Scale» command to enlarge the image to a size of 2048x2048 and download it to your computer.

      – «Inference Steps». The picture we create at the first stage consists of so-called random noise. The program, analyzing our request, reduces this noise step by step and gives us the final version of the image. This option determines the number of steps after which we obtain a sufficiently detailed image. You can try changing it by setting a value in the range from 25 to 50. Increasing the value affects system performance. The optimal ratio of speed and quality of Easy Diffusion is achieved at a value of 25.

      – «Guidance Scale». This option allows us to tell Easy Diffusion to analyze our request more closely so that the final image it produces more accurately matches its description. The optimal value is 7.5. The higher the value, the less creativity Easy Diffusion has.

      – «LoRA» (Low-rank adaptation). LoRA models are used to fine-tune and speed up the large models that we previously downloaded and installed and will use to create our images. There are various LoRA models that are suitable for enhancing some part of the image (for example, eyes) or style (for example, anime). By default, they are not installed on the system. You can download them from the website civitai.com by typing «LoRA» in its search bar. We do not change this option and leave it disabled (empty). However, I recommend that after training, you go to civitai.com and download the models that suit you for further experiments.

      – Seamless Tiling. This option allows you to make a double image seamlessly stitched together from one of your images, that is, make a panoramic image. We do not change this option and leave it disabled (empty).

      – «Output Format». This option specifies the final image format. You can choose three: jpeg, png and webp. I recommend leaving jpeg.

      – «Image Quality». The default value is 75. If you want a higher quality image but need to wait a little longer to create it, set the value to 95.

      We will look at the «Render Settings» functions later in the exercise. For now we leave them unchanged.

      Let’s start creating our first image.

      In order to create our first image from text, in the «Enter Prompt» window we need to write our first text request («Prompt»).

      Let’s do the following:

      – In the «Enter Prompt» window, write the following request: «Artificial Intelligence that will change the world, beautiful cyborg face.» Please note that Easy Diffusion only understands queries in English. If you are not very good at it, please use any online translator available to you.

      – In the