# KNexus Prompt Guide|How to Create Perfect Images Using Prompts

By [KNexus](https://paragraph.com/@knexus) · 2023-08-11

---

Guide to Prompts
----------------

Want to Create Perfect Images with AI but Don't Know How to Use Prompts?

We all know that the key to AI art lies in prompts. When you're using AI art engines like Midjourney or Stable Diffusion to generate images, you might find yourself puzzled when it comes to using the "/:" command to input prompts. You might wonder, how should I describe the image in my mind? This is where the KNexus Prompt Guide comes to your rescue, helping you master the art of using prompts. Of course, we highly encourage you to dive into the world of AI creation with [KNexus](https://knexus.xyz/)!\[Remember to sign up for the Early Bird waitlist on our homepage👀\]

So, let's embark on this learning journey together!

TL;DR:
------

Prompting Techniques
--------------------

When it comes to building effective prompts, there are several techniques and strategies that can help you achieve better results:

1.  **Selecting Keywords**: Keywords play a crucial role in image generation. Choose keywords that are relevant to the content, style, and atmosphere you desire in the image. Strive for specificity and provide detailed information to ensure that the model understands your intent.
    
2.  **Describing Specific Scenes**: To obtain more precise image generation results, provide specific scene descriptions, including the environment, background, lighting, and time of day. These details can help the model better comprehend your expectations.
    
3.  **Experimenting with Styles and Appearances**: If you have specific style preferences, incorporate them into your prompt. Mentioning specific artists, art styles, painting techniques, or photo filters can assist the model in emulating the desired look.
    
4.  **Adjusting Keyword Weights**: You can control the importance of keywords by adjusting their weights. Increasing the weight of certain keywords will make the model more inclined to generate related content. You can try appending a colon and a value to a keyword to adjust its weight, for example: (keyword: 2.0). By fine-tuning the weights, you can obtain more or fewer specific elements in the generated image.
    
5.  **Using Negative Prompts**: If you want to avoid certain content or features, negative prompts can be employed to exclude them. Negative prompts can include objects, styles, or attributes that you wish to avoid, guiding the model away from generating such content.
    
6.  **Iterating and Experimenting**: Image generation is an iterative and experimental process. After generating a few images, assess how well they align with your expectations and fine-tune the prompts as needed. Try different combinations, orders, and weights, and observe their impact on the image results.
    
7.  **Leveraging Model Diversity**: Different temperature values and sampling steps in the generation process can yield images with varying styles, details, and diversity. Experiment with adjusting the temperature and sampling steps to explore different image variations.
    
8.  **Understanding Model Limitations**: Models may have limitations in generating certain objects, details, or complex scenes. Familiarize yourself with the model's limitations and try appropriate prompts and adjustments to address them.
    
9.  **Multiple Attempts and Selecting the Best Results**: Due to the probabilistic nature of image generation, the same prompt may produce different results. Make multiple attempts and select the most satisfactory results.
    
10.  **Continued Learning and Practice**: Prompt construction is a skill that requires continuous learning and practice. By experimenting with different prompts and techniques, you will gradually become familiar with the model's behavior and the features of the generated images, leading to improved results.
    

We hope these techniques will assist you in achieving better image generation outcomes. Good luck with your image generation endeavors! If you have any further questions, feel free to ask.

Structure of a Good Prompt
--------------------------

A good prompt should be detailed, clear, and specific.

The categories of prompts include:

1.  Theme
    
2.  Medium
    
3.  Style
    
4.  Artist Style
    
5.  Website
    
6.  Resolution
    
7.  Details
    
8.  Colors
    
9.  Lighting
    

Next, we'll go through each category of prompts based on Stable Diffusion v1.5. The images are generated with SDXL, 30 steps of DPM++ 2M Karras sampler and image size 1024\*1024.

### Theme

The **theme** is what you want to see in the image. A common issue is that the description of the theme is not specific, precise, or detailed enough.

Let's say we want to generate an image of Audrey Hepburn, no offense to the Angel, we just want to demonstrate how the prompt effect the generated content. A novice might simply write:

> Audrey Hepburn

![](https://storage.googleapis.com/papyrus_images/0f377d8ce8994d30eadd784fba4c4fc425687c91e2b282dd4627fcb939634231.png)

This leaves too much room for imagination. Which scene do you want? What descriptive words can define the scope of the image more precisely? What expression she has? Where is she? What is she doing?

The characteristics of the diffusion model determine that it needs a given content as a reference to associate, reason, and diffuse our ideas. Therefore, we must express what we want accurately.

As an example, let's provide a more specific and precise prompt.

> Audrey Hepburn, lying on the bed, smiling, in black evening dress, knot, looking at the viewer, holding a cup of red wine in glass

![](https://storage.googleapis.com/papyrus_images/a131738d4ecff6fd18742bbd3b27a1207cfc01385de5fb5b6eb2da767546f6bb.png)

### Medium

Medium refers to the materials used to create the artwork. Some examples include illustrations, oil paintings, 3D renderings, and photography. The medium has a strong influence, and using a specific prompt can greatly change the style.

Let's add the prompt **oil** **painting**.

> Audrey Hepburn, lying on the bed, smiling, in black evening dress, knot, looking at the viewer, holding a cup of red wine in glass, **oil painting**

![](https://storage.googleapis.com/papyrus_images/473d6572e5fe7e224c63f12283343da22bbdc1664dc1c63c8ce630f5461e37ff.png)

We can see the desired effect! The image transitions from a photograph to a digital painting.

### Style

**Style** refers to the artistic style of the image, such as Impressionism, Surrealism, Pop Art, etc.

Let's add the prompts **hyperrealistic**, **fantasy**, and **impressionistic** to the prompt.

> Audrey Hepburn, lying on the bed, smiling, in black evening dress, knot, looking at the viewer, holding a cup of red wine in glass, **oil painting**, **hyperrealistic, fantasy, impressionistic**

![](https://storage.googleapis.com/papyrus_images/e982888c8ca6cf97974eb3f80f81a5c73f35dc63c7bc545a87565ad44a3baf2e.png)

### Artist

The name of an artist is a powerful modifier. It allows you to reference specific artists to adjust the style. It is also common to mix the styles of multiple artists.

> Audrey Hepburn, lying on the bed, smiling, in black evening dress, knot, looking at the viewer, holding a cup of red wine in glass, **oil painting**, **hyperrealistic, fantasy, impressionistic**, created by **Paul Cezanne**

![](https://storage.googleapis.com/papyrus_images/e6df96ebb404420e692855697843908078d58e07cf0ebaa41a699e369cfe0e60.png)

### Website

Graphics websites like [Artstation](https://www.artstation.com/) and [Deviant Art](https://www.deviantart.com/) collect various types of images. Using them promptly ensures that the image tends toward those styles.

Let's add **artstation** to the prompt.

> Audrey Hepburn, lying on the bed, smiling, in black evening dress, knot, looking at the viewer, holding a cup of red wine in glass, **oil painting, hyperrealistic, fantasy, surreal**, created by **Paul Cezanne, artstation**

![](https://storage.googleapis.com/papyrus_images/fc5a74bd0cc6938be3621f9dd7bc0516892a785b1f47c7804cf70cf022e5fd14.png)

### Resolution

Resolution represents the clarity and level of detail in the image. Let's use the prompts with **highly detailed** and **sharp focus**.

> Audrey Hepburn, lying on the bed, smiling, in black evening dress, knot, looking at the viewer, holding a cup of red wine in glass, **oil painting**, **hyperrealistic, fantasy, surreal**, created by **Paul Cezanne**, **highly detailed**, **sharp focus**

![](https://storage.googleapis.com/papyrus_images/1bf50d01abc67a0a1a483e23e44f6c4c57a81d13e3b181d059189b0e6f759b14.png)

### Additional Details

Additional details are a way to modify the image. Let's add **modern**, **stunning dresses**, and **cozy themes** to enhance the atmosphere of the image.

> Audrey Hepburn, lying on the bed, smiling, in black evening dress, knot, looking at the viewer, holding a cup of red wine in glass, **oil painting, hyperrealistic, fantasy, surreal,** created by **Paul Cezanne**, **highly detailed, sharp focus, modern, stunning dresses**, and **cozy themes**

![](https://storage.googleapis.com/papyrus_images/9fe0f32352b4f6b730b0c18ea6d11cc2c661edde6e8c8e7e0ea524ba31f548fb.png)

### Colors

You can control the overall color of the image by adding **color prompts**. The colors you specify may appear as tones or appear in objects.

Let's use the prompt **warm-toned** to add some red to the image.

> Audrey Hepburn, lying on the bed, smiling, in black evening dress, knot, looking at the viewer, holding a cup of red wine in glass, **oil painting, hyperrealistic, fantasy, surreal,** created by **Paul Cezanne**, **highly detailed, sharp focus, modern, stunning dresses**, and **cozy themes, warm-toned**

![](https://storage.googleapis.com/papyrus_images/a9da4f039d1259776d7417f7e9eebdc1bef553c4c48c84f79f3dea93ada58016.png)

### Lighting

Any photographer will tell you that lighting is a key factor in creating successful images. Lighting prompts have a significant impact on the appearance of the image. Let's add **cinematic lighting** and **darkness** to the prompt.

> Audrey Hepburn, lying on the bed, smiling, in black evening dress, knot, looking at the viewer, holding a cup of red wine in glass, oil painting, hyperrealistic, fantasy, surreal, created by Paul Cezanne, highly detailed, sharp focus, modern, stunning dresses, and cozy themes, warm-toned, cinematic lighting, darkness

![](https://storage.googleapis.com/papyrus_images/ff41040248d6ae15e449ccb1faed38b4b6e765121c01ff1e9391e78195f38ba8.png)

With these prompts, our prompt becomes more precise and complete.

### Note

_You may have noticed that by adding a few prompts to the theme, the image is already quite good. In building prompts for the diffusion model, it is often not necessary to use many prompts to get good results._

Negative Prompts
----------------

Using negative prompts is another good way to guide the style of the image, but instead of including what you want, you include what you don't want. They don't necessarily have to be objects; they can also be styles and undesired features (e.g., ugly, deformed).

I'll use a generic negative prompt.

> ugly, repetitive, poorly drawn hands, poorly drawn feet, poorly drawn faces, out of frame, extra limbs, distorted, deformed, body out of frame, anatomical errors, watermarks, signatures, cropped, low contrast, underexposed, overexposed, bad artwork, beginner, amateur, distorted faces, blurry, sketchy, grainy

![](https://storage.googleapis.com/papyrus_images/6a69c0c8da5b6385a4509a9a96194b855f9f8624519fe815edaf653073c5354e.png)

Negative prompts help to make the image stand out and prevent it from being mundane.

The Process of Building Good Prompts
------------------------------------

### Iterative Prompt Building

You should approach prompt building as an iterative process. As seen in the earlier sections, simply adding a few prompts to the theme can result in a very good image.

### Using Negative Prompts

Adding negative prompts is part of the iterative process of prompt building. These prompts can be objects or body parts to avoid (e.g., "hands" to hide poorly drawn hands since the v1 model struggles with rendering hands).

Prompting Techniques
--------------------

You can modify the importance of prompts by switching to another prompt in specific sampling steps.

The following syntax applies to KNexus.

### Prompt Weight

(\*This syntax applies to KNexus.\*)

You can adjust the weight of a prompt using syntax `(prompt: factor)`. The `factor` is a value where less than 1 indicates lower weight and greater than 1 indicates higher weight.

For example, let's adjust the weight of the prompt "dog":

> **dog**, autumn in Paris, glamorous, beautiful, ambiance, feeling, mist, smoke, flames, chimneys, rain, moist, pure, puddles, melting, dripping, snow, creek, lush, ice, bridge, forest, roses, flowers, created by Stanley Artgerm Lau, Greg Rutkowski, Thomas Kindkade, Alphonse Mucha, Loish, and Norman Rockwell.

Increasing the weight of "dog" tends to generate more dogs, while decreasing the weight tends to generate fewer dogs. This is not always true for every image, but statistically it holds.

This technique can be applied to theme prompts and all categories, such as style and lighting.

### () and \[\] Syntax

(_This syntax applies to KNexus._)

Another way to adjust the strength of a prompt is by using () and \[\]. (prompt) increases the strength of the prompt by 1.1 times, which is the same as (prompt: 1.1). \[prompt\] decreases the strength by 0.9 times, which is the same as (prompt: 0.9).

You can use multiple parentheses, just like in algebra...the effect is multiplication.

(prompt): 1.1 ((prompt)): 1.21 (((prompt))): 1.33

Similarly, using multiple brackets has the effect of:

\[[prompt](https://chat.openai.com/c/0.9)\]: 0.81 \[\[[prompt](https://chat.openai.com/c/0.9)\]\]: 0.73

### Prompt Blending

(\*This syntax applies to KNexus\*)

You can blend two prompts together. The correct term is **prompt blending**. The syntax for prompt blending is:

> \[prompt1 : prompt2 : factor\]

The `factor` controls when the prompt switches from prompt1 to prompt2. It is a number between 0 and 1.

For example, if I use the following prompt:

> \[Angelina Jolie : Megan Fox : 0.5\] oil painting portrait

![](https://storage.googleapis.com/papyrus_images/497775b25b11dc9c90e1aa3b5fe6b7952445749ef3fb0f5351720f459541a59e.png)

> \[Angelina Jolie : Megan Fox : 0.1\] oil painting portrait

![](https://storage.googleapis.com/papyrus_images/ab9ee20c2b2b48d019b06270c6ad11950b94729d0b7d1449e68abf3aa2886e40.png)

> \[Angelina Jolie : Megan Fox : 0.75\] oil painting portrait

![](https://storage.googleapis.com/papyrus_images/bdedbf4a5d846431e56da0425d5fcfadf416786cdb8a734f3f5fc0f882edf508.png)

for 30 sampling steps.

This means that in steps 1 to 15, the prompt is:

> oil painting portrait of Angelina Jolie

while in steps 16 to 30, the prompt switches to:

> oil painting portrait of Megan Fox

The factor determines when the prompt change happens. It is 15 steps after 30 steps x 0.5 = 15 steps.

Changing the factor has the effect of blending these two actresses to different degrees.

A very important rule with prompt blending is that the **first prompt determines the overall composition**. Early diffusion steps determine the overall composition. Subsequent steps refine the details.

#### Blending Faces

A common use case is creating a new face by blending specific features from actors and actresses. For example, \[Sophie Marceau : Jennifer Connelly : 0.85\], 50 steps are the appearance between the two:

![](https://storage.googleapis.com/papyrus_images/88d5bf6389f9a18d79b2d246462b1692c61e71e8414bde050241de7ddb1f4ced.png)

By carefully selecting the two names and adjusting the factor, we can precisely achieve the desired appearance.

#### Prompt-to-Prompt

With prompt blending, you can achieve similar effects to prompt-to-prompt and generate pairs of highly similar images with variations. Let's try with same parameters like seed and prompt, to check how the prompt-to-prompt works.

> \[Sophie Marceau : Jennifer Connelly : 0.85\] holding an \[pen:glass:0.1\]

![](https://storage.googleapis.com/papyrus_images/ff56ed72c68cf980f25abbb93f1ffa0e5a75e9d366a381cd86163a92378ec8dd.png)

> \[Sophie Marceau : Jennifer Connelly : 0.85\] holding an \[pen:glass:0.9\]

![](https://storage.googleapis.com/papyrus_images/a97f370d5a85a647e0081889cd35f3f253b72659dc018a55ffdd1b8963f20e7d.png)

Careful adjustment of the factor is required. How does it work? The theory behind it is that the overall composition of the image is set early in the diffusion process. Once diffusion gets trapped in a small space, swapping any prompts has little effect on the overall image. It only changes a small part.

Length of Prompts in KNexus
---------------------------

KNexus supports up to 1000 characters for prompts.

Checking Keywords
-----------------

Keywords do not always mean effectiveness. You can check the effectiveness of prompts by using them as standalone prompts. Sometimes the model cannot tell the difference between Charles White and his art, but it knows Andrew Wyeth's painting:

> Andrew Wyeth

![](https://storage.googleapis.com/papyrus_images/49983bbdf2e30f1db785af5a45692b801b9a0ed103d97e8880d9dbef7132d541.png)

> Charles White

![](https://storage.googleapis.com/papyrus_images/5c61b04cbdf99297279c640fc3ca710bcfd13c80968f23ab26d7bc6264b3d4ca.png)

Changing Constraints
--------------------

To excel at prompt building, you need to think like diffusion. The core is an **image sampler** that generates pixel values we, as humans, find reasonable and good. You can even use it without prompts, and it will generate many unrelated images. In technical terms, this is called unconditional or unguided diffusion.

Prompts are a way to guide diffusion into the sampling space that matches your intentions. I mentioned earlier that prompts need to be detailed and specific. This is because detailed prompts narrow down the sampling. Let's look at an example.

> Castle

![](https://storage.googleapis.com/papyrus_images/2f193480e45bb9bb995f196c428c5fe5c2512695136e83871d079a320f44ce76.png)

> Castle, **with a blue sky background**

![](https://storage.googleapis.com/papyrus_images/e27d85d1a0b7e75a691165bd411178bb2479fb8c9604a0e89bebd568aba6487d.png)

> Wide-angle view of a castle, **with a blue sky background**

![](https://storage.googleapis.com/papyrus_images/0f03dd6968257b30ea66d469422711fc83390b361dee51973ac60f17050b7f34.png)

By adding more descriptive prompts to the theme, we narrow down the sampling for castles. In the first example, we ask for **any image of a castle**. Then we ask only for castles with a blue sky background. And finally, we specify it to be a wide-angle photo.

The more specific details you provide in the prompts, the fewer changes in the image.

Association Effects
-------------------

### Attribute Association

There are strong associations between certain attributes. When you specify one attribute, you tend to get another attribute. The diffusion model generates images with unexpected association effects.

Let's say we want to generate a photo of a woman with **blue eyes**.

> A young woman with **blue eyes**, highlights in her hair, sitting outside a restaurant, wearing white clothing, side lighting

![](https://storage.googleapis.com/papyrus_images/e77cb565e344920b1006553b2e088741b0433e828dd1f2b7fcc6058bcca693c9.png)

What if we change it to black eyes?

> A young woman with **black eyes**, highlights in her hair, sitting outside a restaurant, wearing white clothing, side lighting

![](https://storage.googleapis.com/papyrus_images/b84a767b3779afb591059c30dc8a6e59617fe5d71c04e322c56617fdb864f871.png)

Brown Eyes

In the prompts, I didn't specify the race. However, since blue eyes are predominantly found in Europeans, the generated image will be predominantly white. Black eyes are more common across different races, so you will see more diverse racial samples.

Stereotypes and biases are an important topic in AI models.

### Celebrity Name Association

Every prompt has some unexpected associations. This is particularly true for celebrity names. Some actors and actresses have specific poses or wear certain clothes in photos, which are reflected in the training data. If you think about it, model training is just learning through association. If (in the training data) Leonardo DiCaprio always points his fingers at someone, the model will associate the fingers with Leonardo DiCaprio.

When you use Leonardo DiCaprio in prompts, you might expect to see his face. But the pose and clothing of the theme also have an impact. You can study this effect by using just his name as a prompt.

Poses and clothing are part of the overall composition. If you want only his face without the pose, you can use prompt blending to switch his in later sampling steps.

### Artist Name Association

Artist names are particularly prone to association effects.

> digital painting of \[Sophie Marceau : Jennifer Connelly : 0.85\] by Alphonse Mucha

![](https://storage.googleapis.com/papyrus_images/e28623f7e76e113e984f878753565bdcb83eaea3d1de51fdf3670336824369c4.png)

The 19th-century Czech painter Alphonse Mucha is a common occurrence in portrait prompts because the name lends itself to interesting decorations and his style pairs well with digital illustration. However, it often leaves circular or domed patterns in the background. In outdoor environments, they might look unnatural.

Impact of Custom Models
-----------------------

Using custom models is the easiest way to achieve specific styles and is the unique appeal of the diffusion models.

> A young woman with **brown eyes**, highlights in her hair, sitting outside a restaurant, wearing white clothing, side lighting
> 
> In cetusMix\_v4 model:

![](https://storage.googleapis.com/papyrus_images/4d11c81a9a783efecd03c1213af413d1de21fa0a45560d79a71e230d2c331d27.png)

> In Camelmix25D\_v2 model:

![](https://storage.googleapis.com/papyrus_images/20a98e8ea4012fd7eff9869ab60444bfac0fc2e7d7a40b17524106dff3f25c4e.png)

When using models, we need to be aware that the meaning of prompts may change. This is especially true for styles.

Region-Specific Prompts
-----------------------

You can also specify different prompts for different regions of an image.

For example, you can have objects on the left or right side.

**Reference**

[_https://stable-diffusion-art.com/prompt-guide/_](https://stable-diffusion-art.com/prompt-guide/)

\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_\_

### **🧩 About KNexus 🧩**

An AI-driven creative Bazaar, empowering AI enthusiasts to push the limits of creativity with prompts. Safeguard your ideas with lock encryption technology, ensuring rights protection. Generate, create, and purchase innovative creations in this dynamic platform.

[Website](https://knexus.xyz/) | [Twitter](https://twitter.com/KNexus_KNN3) | [Discord](https://discord.com/invite/BabRGjHHWm) | Telegram | [Mirror](https://mirror.xyz/knexus.eth) | [YouTube](https://www.youtube.com/@KNexus-PromptIgnitesInte)

---

*Originally published on [KNexus](https://paragraph.com/@knexus/knexus-prompt-guide-how-to-create-perfect-images-using-prompts)*