AI Drawing] Stabel Diffusion Tutorial, SD Drawing Basics Examples

In the past few days, a couple of students from the AI exchange group have written me privately to say they want to learn more about theSD Mapping(SD is software)Stabel DiffusionIt is an acronym for a very powerfulAI painting tools), because I rarely talk about SD drawing before, and after reading other people's information about SD drawing, I was surprised to find that I didn't know how to get started when I opened the SD interface, or what the parameters did or how to set them.

So, today's article will take an example to give SD mapping newbies some introduction to SD basics, SD interface. Since SD runs on a local computer, there are requirements for computer performance and graphics card:

Key points of computer configuration: graphics card, memory, hard disk, CPU, operating system.

Graphics card: N card (Nvidia Nvida independent graphics card) preferred, efficiency far beyond the set of graphics/AMD/Intel graphics card and CPU rendering, the minimum of 10 series to start with, the experience of a good sense of the 40 series, the minimum of 4G memory, 6G and qualification, on the top of the ceiling; Memory: a minimum of 8G, 16G and qualification, on the top of the ceiling; Hard disk: the best available space has a 500G towards the top of the solid-state is the best.

System Requirements: Supports Win10/Win11/macOS (Apple Silicon only, Intel version of the Mac can not call Radeon graphics) and Linux systems, Apple version of the SD is compatible with a smaller number of plug-ins, the functionality is not as good as Windows and Linux computers.

And for the average person who is not a professional draftsman and just uses AI drawing to make money on projects, investing tens of thousands of dollars in a computer dedicated to SD drawing is not necessary. Because doing projects is not about drawing a lot of diagrams every day like professional diagrammers do.

So what can I do if I want to use SD to draw again? You can use online SD mapping sites, such as MileBlade.com. So, today I'm going to use the Mile High Mile High website as an example to explain it to you:

I. Some basic concepts of SD:

The biggest difference between SD (Stabel Diffusion) and MJ (Midjourney) is that to draw with SD, you first need to select a large model (Checkpoint) in order to draw, and if you want to further draw a more detailed or distinctive picture, you also need a LoRA model.

1. CHECKPOINT

This is the big model necessary for SD drawing, also called the base model, without the big model can't make the drawing. When you open SD drawing, the first thing is to choose the big model, which determines the style type of your final drawing.

In the drawing interface of Mile High Mile High, click "CHECKPOINT", there will be a default model and your own model:

So how do you add the model you want?

● If you have downloaded a large model elsewhere, you can post it to the Mile High website by clicking on the "Publish" - "Models" button at the top of the page.

● Join large models posted by others:

Open the homepage, click "Filter" on the right side of "Models", and select "Checkpoint":

Then pick the model you need.

Or if you need a certain type of large model, such as anime, illustration, landscape, etc., you can directly click on the category to filter out the related large models, and then pick the model you need.

Click on the large model of your choice, e.g. I need "Children's Illustration", click on it and go to the details page.

Click "Join the model library"The large model will be available in the drawing page, and you can select it by clicking "Checkpoint":

2. LoRA model.

This model is not a mandatory model. It is just a model to further adjust the picture effect on the basis of the big model, also called fine-tuning model. In addition, LoRA model is not limited to one, you can use multiple LoRA models together.

LoRA model can be trained by yourself, if you are interested, there is a special training tutorial in Mile High Mile High website, you can check it to learn.

Where do I add the LoRA model in the drawing page? :

Adding LoRA models uploaded by others to "My Model Library" is the same as adding large models, so I won't demonstrate it here.

3. Other parameter parsing:

(1), Vincennes chart:

Get SD to generate images by entering prompt words. (English prompt words are required)

(2), Figure Born Figure.

On top of the existing textual instructions, let Al better understand your needs by providing - a reference picture.

(3) Prompt.

You want to come up with a detailed description of the diagram. (For example..." Long straight black hair, mid 20's, female, wearing a short white tight revealing t-shirt, light blue tight skirt, high heeled black sandals, walking, full frontal view, beach in the background.")

(4) Negative Prompt.

Describe the elements you don't want in the image. (e.g. "ugly", "bad hands", "bad feet", "small eyes" etc.)

(5) Sampler method.

You can choose the sampler, which refers to the different algorithms used in Al's painting. Among them, Euler a is the basic sampler with high universality of effect; DPM++2M Karras sampling method is recommended for faster and better quality; other sampling methods are based on the recommendation of the author's detail page of the model.

(6), the number of iteration steps (Sampling Steps):

Al is required to perform a series of addition and denoising processes while processing the image, and the number of iteration steps indicates the number of times Al simulates the iterative denoising. It is recommended that between 20-30 steps is optimal. The smaller the number of iteration steps, the faster the speed, the lower the quality; on the contrary, the higher the number of steps, the slower the speed, but the higher the quality. 40 steps or more to improve the limited.

(7) High Resolution Restoration.

After setting a lower aspect ratio for image output, adding the HD Repair option further optimizes the image quality. It is essentially the equivalent of doing the graph-to-graph operation again.

(8), redrawing amplitude (Denoising):

Low redraw range means correcting the original image. The higher the redraw magnitude, the more it changes the enlarged image. 0 will not change the original image, 0.3 or less will slightly correct the image based on the original image, more than 0.7 will change the original image greatly, 1 will get a completely different image.

(9), Width and Height.

Resolution adjustment, it is generally recommended to set it around 1024, not too high or too low.

(10), Number of pictures.

Since Al has some instability, you need to try several times to reach a satisfactory result. So, generating multiple pictures can be a good way for us to choose. In Mile High website, you need to upgrade to a paid member to select more than 4 pictures.

(11) Number of elevated word leads.

The higher the value, the more faithful the image generated by Al will be to your cue word. Generally, 7-12 is a safe range of values, the default is 7. Too large or too small a value may lead to abnormal results.

(12) Random Seed.

Clicking on the dice icon means that the picture is randomly generated. Clicking on the green loop arrow continues the creation based on the last generated picture.

II. Example of drawing a picture with SD

Let's use the above mentioned prompt: "Long straight black hair, around 20 years old, female, wearing a white tight short T-shirt with exposed navel, light blue tight skirt, high heeled black sandals, walking with the wind, full frontal view, with the beach in the background" as an example to demonstrate how to use SD to draw your first picture (using Mile High Mile High as a demo).

The step-by-step demonstration is as follows (the LoRA model is not selected for this example):

1. Open the website: https://www.liblib.art/

After registering and logging in, click on "Online Graphics (Web UI)", to access the plotting page.

2. Selection of Large Models.

Click on the "CHECKPOINT" under the large model selection box, you can select the large model, for example, here, I choose the "MAGNUM ORANGE REALISTIC" model :|)

3. Enter the prompt word.

Click on "Text-to-Graphics" and paste the prompt into the prompt text box. Click on "Translate to English" (if the prompt was originally in English, you don't need to click here).

4. Enter the negative cue word:

Just select Default.

5. Select the sampling method and set the number of iteration steps.

By default, the method uses "Euler a" and the number of iteration steps is set to 25.

6、Set the size of the picture

I need a vertical image (9:16), so set the width to 576 and height to 1024, because 576 resolution is too low, so you need to check ""High Resolution Repair"", "Zoom Algorithm" default; adjust the "Redraw Amplitude" to 0.25; "Zoom Magnification" to 2:

7, set the number of images generated

I set it to 2 sheets.

8. Setting the cue word guidance coefficient and the number of random seeds

The default is sufficient, if you need to focus on the prompt word, you can set the value larger.

9、Starting to generate a map

Click on "Start Image Generation" at the top right of the page and the platform will generate the image for us.

The generated images are as follows:

Isn't the generated image not bad? Did you learn it?