1 ‐ Quicktour
Ready to unleash your inner artist? Eikon Diffusion makes AI-powered art accessible to everyone, leveraging the power of the advanced SDXL Stable Diffusion model. In this quick tutorial, we'll show you how easy it is to generate stunning images.
Before we begin:
Make sure you've already set up Eikon Diffusion by following the step-by-step guide: [insert link here].
Let's Get Started!
- Launch Eikon Diffusion: Find the shortcut on your desktop and double-click to open the application.
- Open the Interface: Open your web browser and navigate to the address provided in the application.
-
Time to Get Creative!
- Look for the Prompt text box in the interface. This is where you'll tell Eikon Diffusion what to create.
- Type in: "Picasso style, a squirrel"
- Keep the Negative prompt field at the default value.
- Click the Generate button.
Wait for the image to generate, you should see something like this:
Just below the generated image, in the Image Info section, you'll see a list of parameters used to create it. These parameters are like secret ingredients that control the image generation process. We'll explore each parameter in more detail as we go through this tutorial, but for now, know that you can actually reload these parameters to generate a new image based on the same settings!
Need Inspiration?
If you're not sure how to write the perfect prompt for your desired image style, don't worry! Eikon Diffusion has a handy list of predefined styles to get you started.
- Look for the Prompt text box in the interface. This is where you'll tell Eikon Diffusion what to create.
- Type in: "a squirrel"
- Select in Style: "Picasso"
- Click the Generate button.
That's it! Eikon Diffusion will now use its powerful AI to interpret your prompt and generate a unique image inspired by Picasso's style, featuring a charming squirrel.
Want to explore different ideas simultaneously? Eikon Diffusion lets you generate multiple images from different prompts at the same time!
How to Do It:
One Prompt Per Line: In the Prompt text box, simply type each prompt on a separate line.
For example:
Picasso style, a squirrel
Picasso style, a cat
Picasso style, a dog
Generate: Click the Generate button, and Eikon Diffusion will create a separate image for each prompt.
| Image 1 | Image 2 | Image 3 |
|---|---|---|
|
 |
 |
Important Note: For best results when applying different styles to each prompt, we recommend adding the style directly to each individual prompt rather than using the predefined styles checkbox. This gives you more precise control over the style of each generated image.
Benefits of Generating Multiple Images:
- Explore Variations: Quickly see how different prompts translate into visuals.
- Save Time: Generate multiple ideas without having to run separate generations.
- Spark Creativity: Unexpected combinations can lead to new and exciting concepts.
Eikon Diffusion's batch generation feature lets you easily create multiple variations of a single image prompt.
How to Use Batch Generation:
- Enter Your Prompts: Type your prompt(s) as described in the previous sections (either a single prompt or multiple prompts on separate lines).
- Adjust the parameter Number of output images (per prompt) in the Eikon Diffusion interface. This setting determines how many images will be generated for each prompt.
- Generate: Click the Generate button. Eikon Diffusion will then create the specified number of images for each prompt.
Benefits of Batch Generation:
- Explore Variations: Generate multiple versions of an image with slight differences in composition, style, or details.
- Save Time: Avoid repeatedly clicking the "Generate" button for each image.
- Analyze Results: Easily compare and contrast different generated images side-by-side.
Not Quite What You Imagined?
Even with the most powerful AI, sometimes the first result isn't quite what you envisioned. That's where Samplers come in!
Think of Samplers like different brushstrokes an artist might use. Each one creates a unique texture and feel in the final painting.
Here are a few popular Samplers to get you started:
-
Euler: This eager beginner artist is fast and full of enthusiasm, but might need a little refining.
-
DPM++ 2M Karras: A skilled and experienced artist who strikes a great balance between speed and quality, delivering solid and enjoyable results.
-
DDIM: This master artist takes their time and pays meticulous attention to detail, creating highly polished and impressive pieces.
Want to learn more about each Sampler? Check out this link for detailed information: [insert link here]
Let's Test It Out!
Let's see how the DPM++ 2M Karras Sampler performs with our original prompt.
-
Keep It Consistent: To make a fair comparison, we need to use the same starting point for both images. You can either:
- Reload Parameters: Click the button to reload the parameters from the last generated image.
- Manually Enter the Seed: In the interface, uncheck the Randomize seed option and in the Seed field, enter the seed number you see listed in the Image Info section of the previous image.
- In the interface, select "DPM++ 2M Karras" from the Sampler dropdown menu.
Now, generate a new image and observe the differences!
You've got your prompt ready and your sampler chosen, but what about the size of your image?
Sweet Spot: SDXL diffusion models is built to create stunning images that are 1024x1024 pixels. This resolution allows the model to really shine, showcasing its full creative potential and generating the most detailed and vibrant results.
Going Bigger: Want to push the boundaries? You can generate images up to 1280x1280 pixels with very little loss in quality.
Scaling Down: It's also possible to create smaller images, starting from 512px. However, for the best possible results and to fully leverage the power of model, we recommend generating at least a 1024x1024 image first and then resizing it down if you need something smaller.
Choosing Your Image Size
Eikon Diffusion gives you flexibility in choosing the perfect resolution for your image.
Pixel Perfect: Want complete control? Select the Pixel option and use the Width and Height selectors to enter your desired dimensions in pixels.
Aspect Ratio Magic: Prefer a specific look and feel? Selete the Aspect ratio option and choose from our handy list of aspect ratio templates! Each template is pre-set to common image proportions, making it easy to get the right look.
Aspect Ratio Templates:
Below you'll find a list of our available aspect ratio templates, along with their corresponding resolutions:
| Template | Width | Height |
|---|---|---|
| Instagram (1:1) | 1024 | 1024 |
| 35mm film / Landscape (3:2) | 1024 | 680 |
| 35mm film / Portrait (2:3) | 680 | 1024 |
| CRT Monitor / Landscape (5:4) | 1280 | 1024 |
| CRT Monitor / Portrait (4:5) | 1024 | 1280 |
| CRT Monitor / Landscape (4:3) | 1024 | 768 |
| CRT Monitor / Portrait (3:4) | 768 | 1024 |
| Widescreen TV / Landscape (16:9) | 1024 | 576 |
| Widescreen TV / Portrait (9:16) | 576 | 1024 |
| Widescreen Monitor / Landscape (16:10) | 1024 | 640 |
| Widescreen Monitor / Portrait (10:16) | 640 | 1024 |
| Cinemascope (2.39:1) | 1024 | 424 |
| Widescreen Movie (1.85:1) | 1024 | 552 |
| Academy Movie (1.37:1) | 1024 | 744 |
| Sheet-print (A-series) / Landscape (297:210) | 1024 | 720 |
| Sheet-print (A-series) / Portrait (210:297) | 720 | 1024 |
Level Up Your Images: Two Powerful New Upscalers!
Eikon Diffusion is getting even better with the addition of two powerful new upscalers: SAFMN++ and ScaleCrafter. We'll explore both in detail, but let's start with SAFMN++. For ScaleCrafter will get its own dedicated section later in the tutorial.
What it does: SAFMN++ takes images, even those saved in formats like PNG, and intelligently boosts their resolution, making them sharper and more detailed. Think of it as giving your images a high-definition makeover!
Why it's special: SAFMN++ is designed for real-time performance, meaning it can upscale images quickly without any lag. It's particularly good at handling images compressed with formats like AVIF, which are known for their superior quality compared to older formats like JPEG.
How it works: SAFMN++ uses a clever technique called "spatially-adaptive feature modulation" to analyze and enhance the important details in your images. This results in incredibly natural-looking upscaled images that look like they were originally captured at a higher resolution.
Let's Upscale!
Remember that adorable squirrel we generated earlier? Let's give it a high-resolution makeover with SAFMN++!
- Access the Upscaling tab.
- Scroll down to the bottom of the SAFMN section.
- Select the Use SAFMN++ option.
- Under Upscale select the 4x option.
- Click the Generate button.
And here's the result!
| 1024px | 4096px - 4k (upscaled image) |
|---|---|
|
|
| Download | Download |
Take a close look at both images. Notice how the upscaling might affect the boldness of the lines, the definition of the geometric shapes, and the overall impact of the composition. SAFMN++ is designed to intelligently enhance details, so see if you can spot subtle improvements in the clarity and expression of our Picasso squirrel!
Let's say you want to create an image of a Persian cat playing with a ball of yarn. You describe it in a text:
"A Persian cat playing with a ball of yarn".
For a computer to understand this description and create the image, we need to translate your words into something it can understand. That's where the tokenizer and text encoder come in.
- Tokenizer: Cutting the sentence into pieces:
The tokenizer is like a word cutter. It takes your sentence "A Persian cat playing with a ball of yarn" and divides it into smaller pieces called tokens: ["A", "Persian", "cat", "playing", "with", "a", "ball", "of", "yarn"].
- Text Encoder: Transforming words into numbers:
Now, each token needs to be transformed into something the computer can understand: numbers! The text encoder does this. It associates each token with a numerical vector, which is like a mathematical representation of the word.
Imagine that "cat" is represented by the vector [0.2, 0.8, 0.1], "Persian" by [0.9, 0.1, 0.6], and so on. These vectors capture the meaning of each word.
Stable Diffusion XL, a powerful model for generating images, has a limit of 77 tokens per prompt. This means your description can't be too long. Also, we can't emphasize or de-emphasize parts of the text, like "ball of yarn", to make the model give more or less importance to that concept.
We know that the 77-token limit and the inability to emphasize certain words can be frustrating. That's why Eikon Diffusion offers two powerful features to overcome these limitations: Prompt weighting and Long prompts.
To make things simpler for you, we've combined these techniques under the umbrella term "Prompt Methods". When generating an image, you'll simply choose the embedding generation method that best suits your needs. Here are your options:
- Normal: This is the standard Stable Diffusion XL method. It doesn't allow for text weighting and is limited to 77 tokens.
- Long Prompt Weighting: This method, originally developed by Andrew Zhu and implemented in Eikon Diffusion, allows you to assign weights to different parts of your text, emphasizing or de-emphasizing specific concepts. It also removes the 77-token limit, giving you more freedom to craft detailed descriptions.
- Compel: This is another text prompt weighting and combining library that offers similar functionality to Long Prompt Weighting. It uses a different and potentially simpler syntax for weighting, and may produce slightly different results due to its unique implementation.
By offering these diverse Prompt Methods, Eikon Diffusion empowers you to create more nuanced and complex images, pushing the boundaries of what's possible with AI-generated art.
Now, let's see how methods 2 and 3 work:
The Prompt Weighting method allows us to control the importance of different parts of our prompt. Think of it like highlighting or underlining words in a text to emphasize or de-emphasize them.
- To emphasize parts of the prompt, we use parentheses: (ball of yarn)
- To de-emphasize parts of the prompt, we use square brackets: [ball of yarn]
- To emphasize by defining a specific weight, we use the following format: (ball of yarn:1.3)
Here, "1.3" represents the weight we assign to "ball of yarn". A weight higher than 1.0 emphasizes the term, while a weight lower than 1.0 de-emphasizes it.
Prompt weighting equivalents:
| ball of yarn == (ball of yarn:1.0) (no emphasis) | (ball of yarn) == (ball of yarn:1.21) (slightly emphasized) | ((ball of yarn)) == (ball of yarn:1.5) (more emphasized) | [ball of yarn] == (ball of yarn:0.91) (slightly de-emphasized) | (ball of yarn:0) (highly de-emphasized) |
|---|---|---|---|---|
 |
 |
 |
 |
 |
It's time to try it out!
Test it yourself by adjusting these settings in the interface:
- Enter the Prompt with the desired weight: For example, you could try "Persian cat playing with a ball of yarn:1.5".
- Remove the Negative prompt: For this exercise, leave the negative prompt field empty.
- Change the Prompt method value to "Long prompt weighted": This tells the system to use the weighting you've defined.
- Click the Generate button: Now, watch as the AI creates an image based on your weighted prompt!
Experiment with different weights and see how they affect the generated image. You'll be amazed at the level of control you can achieve!
Now that you've seen how to use it, let's see how it performs with a longer prompt.
Let's use an example provided by Andrew Zhu and compare the results of the long prompt applied in the normal method and in the Long Prompt Weighted method. This will help us understand the benefits of using Long Prompt Weighting for more complex prompts.
- Put these prompts in the Prompt and Negative prompt field:
Prompt:
A whimsical and creative image depicting a hybrid creature that is a mix of a waffle and a hippopotamus. This imaginative creature features the distinctive, bulky body of a hippo, but with a texture and appearance resembling a golden-brown, crispy waffle. The creature might have elements like waffle squares across its skin and a syrup-like sheen. It's set in a surreal environment that playfully combines a natural water habitat of a hippo with elements of a breakfast table setting, possibly including oversized utensils or plates in the background. The image should evoke a sense of playful absurdity and culinary fantasy.
Negative prompt:
skin spots, acnes, skin blemishes, age spot, (ugly:1.2), (duplicate:1.2), (morbid:1.21), (mutilated:1.2), (tranny:1.2), mutated hands, (poorly drawn hands:1.5), blurry, (bad anatomy:1.2), (bad proportions:1.3), extra limbs, (disfigured:1.2), (missing arms:1.2), (extra legs:1.2), (fused fingers:1.5), (too many fingers:1.5), (unclear eyes:1.2), lowers, missing fingers, extra digit, bad hands, (extra arms and legs), (worst quality:2), (low quality:2), (normal quality:2), lowres, ((monochrome)), ((grayscale))
- Change the Prompt method value to "Long prompt weighted".
- Use this Seed number: 1679219438.
- Make sure the Randomize Seed option is unchecked.
- Keep sampler in "DPM++ 2M Karras".
- Click the Generate button.
| Normal Method | Long Prompt Weighted Method |
|---|---|
 |
 |
By comparing the results generated by both methods, we can see how Long Prompt Weighted allows for finer control over the image generation process, especially when dealing with longer and more detailed prompts.
In the previous section, we explored Long Prompt Weighted. Now, let's dive into another powerful feature for controlling your image generation: Compel.
Compel offers a unique approach to prompt weighting, using a syntax that might feel more intuitive to some users. Instead of parentheses and numerical weights, Compel leverages plus (+) and minus (-) signs to directly adjust the importance of words or phrases within your prompt.
How Compel Works:
Think of Compel's + and - symbols as dials that fine-tune the "attention" the AI model pays to specific elements in your prompt.
- + (Plus): Increases the weight of a word or phrase, making it more prominent in the generated image.
- - (Minus): Decreases the weight, making the element less prominent or even subtly downplaying its presence.
Syntax Breakdown:
Compel's syntax is quite flexible, allowing you to apply weights to single words, phrases, or even entire sections of your prompt.
Here are some examples:
-
Single words without parentheses:
- A Persian cat playing with a ball of yarn+ (emphasizes "yarn")
-
Single or multiple words with parentheses:
-
A Persian cat playing with a ball of (yarn)+ (emphasizes "yarn")
-
A Persian cat playing with a ball of (yarn)- (de-emphasizes "yarn")
-
A Persian cat playing with a (ball of yarn)+ (emphasizes "ball of yarn")
-
A Persian cat playing with a (ball of yarn)- (de-emphasizes "ball of yarn")
-
Amplifying the effect:
A Persian cat playing with a (ball of yarn)++ (strongly emphasizes "ball of yarn")
Nesting for even stronger emphasis:
A Persian cat playing with a (ball of yarn+)++ (effectively gives "yarn" a +++ weighting)
Alternative Syntax:
You can also use a more direct numerical approach within parentheses:
(ball of yarn)1.1 to emphasize "ball of yarn"
(ball of yarn)0.6 to de-emphasize "ball of yarn"
Understanding the Weight Values:
Compel uses a simple scaling system:
- +: corresponds to a weight of 1.1 (a slight increase in attention)
- ++: corresponds to 1.1 squared (1.21, a stronger increase)
- +++: corresponds to 1.1 cubed (1.331, an even stronger increase), and so on.
Similarly:
- -: corresponds to a weight of 0.9 (a slight decrease in attention)
- --: corresponds to 0.9 squared (0.81, a stronger decrease)
Now let's see how Compel performs in practice!
Test it yourself by adjusting these settings in the interface:
- Enter the Prompt with the desired weight: For example, you could try "Persian cat playing with a (ball of yarn)++".
- Remove the Negative prompt: For this exercise, leave the negative prompt field empty.
- Change the Prompt method value to "Long prompt weighted": This tells the system to use the weighting you've defined.
- Click the Generate button: Now, watch as the AI creates an image based on your weighted prompt!
| ball of yarn+ (emphasizes "yarn") | (ball of yarn)+ (emphasizes "ball of yarn") | (ball of yarn)++ (strongly emphasizes "ball of yarn") | (ball of yarn)-- (de-emphasizes "ball of yarn") | (ball of yarn)0.1 (strongly de-emphasizes "ball of yarn") |
|---|---|---|---|---|
 |
 |
 |
 |
 |
Experiment and Explore!
Feel free to experiment with different combinations of + and - symbols in your prompts. The beauty of Compel lies in its flexibility and the ability to fine-tune your creative vision.
Remember:
Compel offers a different flavor of prompt weighting compared to Long Prompt Weighted. Try both methods and see which one resonates best with your workflow and artistic goals.
Now that you've seen how to use it, let's see how Compel performs with a longer prompt.
Let's use the same example before with the same settings, but changing only the negative prompt.
- Put these prompts in the "Prompt" and "Negative prompt" field:
Prompt:
A whimsical and creative image depicting a hybrid creature that is a mix of a waffle and a hippopotamus. This imaginative creature features the distinctive, bulky body of a hippo, but with a texture and appearance resembling a golden-brown, crispy waffle. The creature might have elements like waffle squares across its skin and a syrup-like sheen. It's set in a surreal environment that playfully combines a natural water habitat of a hippo with elements of a breakfast table setting, possibly including oversized utensils or plates in the background. The image should evoke a sense of playful absurdity and culinary fantasy.
Negative prompt:
skin spots, acnes, skin blemishes, age spot, (ugly)1.2, (duplicate)1.2, (morbid)1.21, (mutilated)1.2, (tranny)1.2, mutated hands, (poorly drawn hands)1.5, blurry, (bad anatomy)1.2, (bad proportions)1.3, extra limbs, (disfigured)1.2, (missing arms)1.2, (extra legs)1.2, (fused fingers)1.5, (too many fingers)1.5, (unclear eyes)1.2, lowers, bad hands, missing fingers, extra digit, (extra arms and legs), (worst quality)2, (low quality)2, (normal quality)2, lowres, ((monochrome)), ((grayscale))
- Change the Prompt method value to "Compel".
- Use this Seed number: 1679219438.
- Make sure the Randomize Seed option is unchecked.
- Keep sampler in "DPM++ 2M Karras".
- Click the Generate button.
| Normal Method | Compel Method |
|---|---|
|
 |
You'll notice a setting called "CLIP skip" in Eikon Diffusion. This setting plays a crucial role in how deeply the AI model understands your text prompt.
What is CLIP?
Think of CLIP as a brilliant translator. It takes your words and transforms them into a language that Stable Diffusion can comprehend.
How CLIP skip Works:
CLIP is like a multi-layered cake, with each layer refining its understanding of your prompt.
- Larger CLIP skip: Imagine stopping the translator after a few layers. This might lead to a more abstract or unexpected interpretation of your prompt.
- Smaller CLIP skip: Allowing CLIP to process your prompt through more layers results in a more precise and detailed understanding.
Does CLIP skip Matter for SDXL Model?
SDXL, the latest generation of Stable Diffusion models, is trained using a specific CLIP layer. This means that changing the CLIP skip setting won't affect the output of SDXL models.
Why Keep CLIP skip in Eikon Diffusion?
While CLIP skip doesn't affect SDXL models, we've kept this feature in Eikon Diffusion because some users have reported interesting variations in image generation when using different CLIP skip values with certain Stable Diffusion models.
Note: CLIP skip's effect varies depending on the Prompt Method used.
Let's illustrate this with a practical example.
- Look for the Prompt text box in the interface. This is where you'll tell Eikon Diffusion what to create.
- Type in: "Picasso style, a squirrel"
- Select in Prompt method: "Normal"
- In the CLIP skip field, enter: 0,
- Select in Sampler: "DPM++ 2M Karras"
- In the Seed field, enter: "1176128203"
- Make sure the Randomize Seed option is unchecked.
- Click the Generate button.
Comparison between different CLIP skip values
| value = 0 (disabled) | value = 1 | value = 2 | value = 3 |
|---|---|---|---|
 |
 |
 |
 |
Feel free to experiment and see how different CLIP skip values influence your image generation!
Now that you understand how Eikon Diffusion harnesses the power of SDXL to bring your ideas to life, let's explore how to fine-tune the image generation process.
Think of the guidance scale as the reins you hold over your AI artist. It controls how closely the generated image adheres to your text prompt.
-
Lower Guidance Scale (e.g., 5-7): Imagine giving your AI artist a loose suggestion rather than a strict instruction. A lower guidance scale allows for more creative freedom, resulting in images that might be inspired by your prompt but take unexpected and interesting turns.
-
Higher Guidance Scale (e.g., 8-15): Here, you're giving your AI artist a more detailed blueprint. A higher guidance scale ensures the generated image closely matches your prompt, but be careful! Going too high can sometimes lead to artifacts or unnatural elements in the image.
Putting Guidance Scale into Practice
Let's see how different guidance scales affect the final image. We'll generate a new image using a different prompt to illustrate the concept.
Resetting to Defaults:
Before we begin, let's reset Eikon Diffusion to its default settings. This ensures we're starting with a clean slate. Select "Default" profile and click the "Load profile" button.
Tip: If you experiment with various settings and encounter difficulties achieving the desired results, remember that you can always restore the default settings by clicking the "Load profile" button. It will always load the last loaded profile. If you're using Eikon Diffusion for the first time, your first profile will be "Default". If it's not already selected, simply choose it before clicking "Load profile".
Now, let's explore how different guidance scales shape our new image!
- In the Prompt field, type: "portrait photo of an old warrior chief".
- Select "DPM++ 2M Karras" from the Sampler dropdown menu.
- In the Seed field, enter: 6.
- Make sure the Randomize Seed option is unchecked.
- Our Guidance scale is already set to the default value of 7, which is a good starting point.
- Click the "Generate" button and watch as was generated!
| Generated image - Guidance Scale 7.0 |
|---|
 |
Comparison between different scales
| Guidance Scale 2.0 | Guidance Scale 5.0 | Guidance Scale 7.0 | Guidance Scale 10.0 | Guidance Scale 15.0 |
|---|---|---|---|---|
 |
 |
|
 |
 |
Finding the Sweet Spot Experiment with different guidance scale values to find the perfect balance between creative exploration and precise control.
Tip: Start with a moderate guidance scale (around 7-9) and adjust based on your desired level of control and the complexity of your prompt.
We've explored how Guidance Scale influences the overall adherence to your prompt. Now, let's delve into Guidance Threshold, which directly controls how strongly the Guidance Scale is applied during image generation.
Think of it like this:
- Higher Guidance Threshold (closer to 1.0): Imagine focusing a microscope. A higher Guidance Threshold ensures the Guidance Scale is applied throughout the entire image generation process, resulting in sharper details and a clearer representation of your prompt.
- Lower Guidance Threshold (closer to 0.1): This is like looking through a blurry lens. A lower Guidance Threshold reduces the impact of your chosen Guidance Scale, potentially leading to a less detailed and less accurate representation of your prompt.
Values below 1.0, down to 0.5, can further reduce inference time with minimal impact on image quality.
Comparison between different values
| Guidance Threshold 0.1 | Guidance Threshold 0.2 |
|---|---|
 |
 |
| Guidance Threshold 0.3 | Guidance Threshold 0.4 |
|---|---|
 |
 |
| Guidance Threshold 0.5 | Guidance Threshold 0.6 |
|---|---|
 |
 |
| Guidance Threshold 0.7 | Guidance Threshold 0.8 |
|---|---|
 |
 |
| Guidance Threshold 0.9 | Guidance Threshold 1.0 |
|---|---|
 |
 |
We've already explored how CFG Scale and Guidance Threshold influence image generation in Eikon Diffusion, guiding the AI like an artist following instructions. Now, let's delve into an even more powerful feature: Perturbed-Attention Guidance (PAG).
Imagine you're sculpting a statue. You start with a rough block of marble and gradually chip away pieces to reveal the final form. PAG works in a similar way, but instead of a chisel, we use the AI's "attention."
The AI's "attention" is like its ability to focus on specific parts of the image and understand how they relate to each other. It's as if the AI is looking at different parts of the image and trying to figure out how they fit together.
PAG temporarily "perturbs" this attention, making it a bit more "blurry." It's like putting a light veil over the AI's eyes, making it see the image less sharply.
But why do this?
This controlled "disruption" forces the AI to work harder to reconstruct the image. By reconstructing the image from these "blurred" parts, PAG helps generate sharper results with richer textures and a more realistic overall appearance.
Benefits of PAG:
- More detailed images: PAG helps capture subtle nuances and details that might be lost with traditional methods.
- Enhanced realism: Images generated with PAG tend to have a more natural and convincing look.
- Greater creative control: PAG offers users more options to refine and personalize their creations.
Putting PAG into Practice
Let's see PAG in action in the image we generated previously.
- In the PAG section of the interface, check the Use PAG checkbox.
- Click the Generate button again and compare the results.
| Without PAG | With PAG |
|---|---|
|
 |
Just like CFG Guidance scale, you can control the intensity of PAG's influence changing the parameter Guidance scale.
Think of Guidance scale as the volume knob for PAG.
- Higher guidance scale (e.g., 5-7): Imagine turning up the volume on PAG. This leads to images with more defined structures and fewer visual glitches. However, just like with in CFG, excessively high values can result in smoother textures and slight saturation in the images.
- Lower guidance scale (e.g., 1-3): This is like turning down the volume. It allows for more creative freedom, potentially leading to more abstract or dreamlike results.
By default the Eikon diffusion uses guidance scale in 3.0, which strikes a good balance for most use cases. This was the value applied in previous image.
Let's generate image again, using values 2.0 and 5.0. Then, compare the results.
| Guidance scale 2.0 | Guidance scale 3.0 | Guidance scale 5.0 |
|---|---|---|
 |
 |
 |
 |
 |
 |
We can see that scale 2.0 outperformed scale 3.0. It preserved some details from the original image, added details to the shoulders, and avoided the incomplete mustache seen in the previous image. The 5.0 scale had good detail, but the feathers on the headdress lost a lot of detail and the image was oversaturated. Feel free to experiment with different values to find the sweet spot for your desired level of detail and realism.
Like Guidance threshold for CFG scale, the Adaptive scale in PAG controls how much of the chosen guidance scale is applied throughout the image generation process.
That is, a value of 0.5 means the guidance scale will be applied in 50% of the generation steps. Higher values increase the influence of guidance, while lower values decrease it. This allows for more nuanced control over the balance between adhering to your prompt and allowing for creative variations.
Let's try 0.5 value:
- Change Adaptive scale value to 0.5 and click on "Generate" button again!
| Without adaptive scale | With adaptive scale |
|---|---|
|
 |
As is evident, the difference was minimal. Lower adaptive scale values demonstrate a tendency to preserve more details from the original image.
The power of PAG doesn't stop at simply turning it on or off. Eikon Diffusion gives you even finer control by allowing you to apply PAG to specific layers within the model's architecture.
Think of the model as a complex machine with many interconnected gears. Each layer represents a different stage of processing, refining the image as it moves through the system.
By default, PAG focuses on the "mid" blocks – the layers responsible for shaping the core structure and details of the image. However, you can change Layers blocks field to target other layers, tailoring PAG's influence to achieve specific effects.
Let's generate two more examples by changing the layers, first we will generate an image using the layers "down.blocks.2 and up.blocks.1.attn.0" and then another image with only the layer "down.blocks.2"
Set the Adaptive scale value back to 0 and change the values of the Layers block parameter to the values in the image below and compare the different results
| down.blocks.2 and up.blocks.1.attn.0 | down.blocks.2 |
|---|---|
 |
 |
 |
 |
Pretty incredible, right? By playing around with different layer combinations, you can unlock unique styles and highlight specific details in your images.
Think of Stable Diffusion as a talented artist who starts with a blurry canvas and gradually reveals a stunning image. This process of removing noise and sharpening details is handled by a special part of Stable Diffusion called the UNet. Even the most skilled artist sometimes needs a little help to achieve the perfect balance of detail and coherence.
That's where FreeU comes in! Think of FreeU as a helpful art director who guides the artist, ensuring that the final masterpiece is both detailed and harmonious.
- How FreeU Works:
FreeU fine-tunes the artist's process by adjusting the influence of two key elements:
- Skip Connections: These are like special shortcuts the artist uses to add fine details, like adding delicate brushstrokes or highlights.
- Backbone Features: These are the core elements of the painting, like the main shapes, colors, and composition.
FreeU helps the artist find the perfect balance between these elements, preventing the image from becoming too cluttered or losing its overall coherence.
Using FreeU in Your Workflow:
To unlock the power of FreeU, simply check the "Use FreeU" option in the FreeU section of Eikon Diffusion's generation tab.
You can easily adjust the balance between detail and coherence by tweaking the S1 scale, S2 scale, B1 scale, and B2 scale settings. Think of these settings like adjusting the artist's brushstrokes – each setting influences how the artist applies detail and refines the overall image.
Now, let's dive into the settings that let you fine-tune FreeU's magic:
S1 and S2 (Skip Connection Scales): Think of these as filters that shape the details generated by Stable Diffusion. They range from 0.00 to 1.0, with 1.0 representing the original image without any FreeU adjustments.
Lowering these values modifies the image's structure and details. S1 primarily handles broader details like the image's pose and overall structure, while S2 focuses on refining finer details.
B1 and B2 (Backbone Feature Scales): These scales amplify the details produced by S1 and S2, adding or reinforcing those final touches.
In Eikon Diffusion, we've made it easy to experiment with FreeU by allowing incremental adjustments of 0.01 for each scale.
Starting Points and Experimentation
The default values are:
- S1 = 0.9
- S2 = 0.2
- B1 = 1.3
- B2 = 1.4
These are good starting points, but feel free to explore! We recommend beginning with both S1, S2, B1 and B2 set to 1.0. Then, experiment by gradually decreasing S1 and S2 while slightly increasing B1 and B2.
Just be careful with values above 1.2 for B2 can sometimes lead to oversaturated images. These numbers are like specific brushstroke settings that control the balance between detail and coherence.
A Practical Example
Back to our old warrior image. We want to preserve some of the raw, gritty details from earlier stages of the generation process, while still benefiting from FreeU’s refinement.
S1 and S2 Refinement: We'll start by keeping S1 value in 1.0 and lowering the S2 scale to 0.65. This will retain some of the original details that might have been smoothed out in the final image.
Amplify with B1 and B2: Next, we'll give those details a subtle boost by increasing B1 and B2 by 0.05 each. This will enhance the preserved details without overdoing it.
Generate and Compare:
Now, let's generate two versions of our warrior image: one with PAG enabled and one without. This will allow us to see how FreeU interacts with PAG and how it affects the final result.
| Original Image | FreeU without PAG | FreeU with PAG |
|---|---|---|
|
 |
 |
Observing the Results
By comparing these two images, you'll gain a better understanding of how FreeU can be used to fine-tune the balance between detail and coherence in your Stable Diffusion creations.
You'll notice that enabling PAG can sometimes lead to a trade-off. In our warrior example, PAG might have smoothed out some details, like the paint on the face, while sharpening others, like the tips of the feathers.
Remember, experimentation is key! Play around with different settings and observe how they affect your results. You'll soon develop a feel for how to achieve the perfect balance of detail and coherence in your Stable Diffusion masterpieces.
Imagine a vast library filled with countless recipes for every imaginable image. This library is called the "latent space," and each recipe is a unique combination of ingredients that represent different aspects of an image – colors, shapes, textures, and more.
When you give a text prompt to an AI image generator like Stable Diffusion, it's like asking the librarian for a specific recipe. The AI then uses this recipe to assemble the ingredients and create a visual masterpiece.
Eikon Diffusion gives you the power to peek into this library and adjust the ingredients directly. Want to make the colors more vibrant? Tweak a few settings. Need to sharpen the details? You can do that.
Let's explore some of the amazing things you can do with Eikon Diffusion:
1. Outlier Removal & Detail Enhancement: Crafting a Cohesive Vision
Imagine your image is like a bustling city. Sometimes, a few unexpected elements might pop up – a rogue building, a misplaced tree, or a splash of color that doesn't quite fit.
Outlier removal is like having a city planner who tidies things up. It identifies and removes those unexpected elements, ensuring your image has a cohesive and balanced look.
Think of it like sharpening those details, bringing your images into focus and revealing the intricate beauty within.
2. Color Balancing: Fine-Tuning the Palette
Have you noticed how some AI-generated images tend to lean towards yellows? It's like the AI has a slight preference for sunshine!
SDXL exhibits a noticeable tendency towards color bias often pushing values beyond the acceptable range (see left image). This issue can be effectively addressed by centering the values, ensuring they fall within the desired boundaries (right image).
| Outside the boundaries | Centered within the boundaries |
 |
 |
| Credit: Timothy Alexis Vass | |
Eikon Diffusion lets you adjust this natural tendency. Think of it like adjusting the balance of colors in your image. You can shift the palette away from that yellow bias, ensuring your images have a more balanced and natural look.
Color balancing is like adjusting the volume of each instrument, ensuring all the colors work together in perfect harmony.
3. Tensor Maximizing: Unleashing the Full Potential
Imagine your image is like a painting. Sometimes, the artist might be using a limited palette, not utilizing the full range of colors available.
Tensor maximizing is like encouraging the artist to explore the full spectrum of colors, unleashing the full vibrancy and detail of the painting.
Putting Theory into Practice
Now that you understand the concepts behind latent space refinement, it's time to see it in action.
- Head over to the Generation tab in Eikon Diffusion. You'll find a section labeled Latent space refinement.
- Make sure the checkbox next to Use latent space refinement is ticked.
-
For now, leave the default values for each parameter as they are. These settings are carefully chosen to provide a good starting point for most images.
-
Click the "Generate" button and watch as works its.
| Original Image | Adjusted w/FreeU and without PAG | Adjusted w/FreeU and with PAG |
|---|---|---|
|
 |
 |
Compare the newly generated images with the one you created without latent space refinement. Do you notice any improvements in detail, color balance, or overall quality?
Tip: If you're increasing the value of one parameter and notice an unexpected change in your image, try proportionally increasing the other parameters as well. Conversely, if you're decreasing a value, consider decreasing the others proportionally.
Even with the power of latent space refinement, sometimes you might need a few extra tweaks to achieve the perfect look. That's where Eikon Diffusion's post-processing tools come in handy.
Fine-Tuning Your Masterpiece:
If you weren't able to achieve the desired color and luminance balance during the generation process, don't worry! In the "Post-processing" tab offers a range of parameters to make those final adjustments.
There, you can fine-tune the RGB color balance, adjust gamma, contrast, brightness, and even apply various filters to enhance your image.
A few subtle adjustments can make all the difference in bringing your vision to life.
Lets see how it works:
- Navigate to the Post-processing tab. Make sure the checkbox next to Use image filters is ticked.
- In the "Other adjustments" section, you'll find parameters for fine-tuning color and luminance. Let's apply the following settings as a starting point:
- Red: 1.1
- Green: 1.05
- Blue: 1.15
- Gamma: 1.3
- Auto Contrast Low: 5
- Color: 1
Click the "Generate" button again to apply these post-processing adjustments. Compare the new image with the previous version. Do you notice any improvements in color balance, contrast, or overall visual appeal?
| Original Image | Adjusted w/FreeU and without PAG | Adjusted w/FreeU and with PAG |
|---|---|---|
|
 |
 |
Explore the Possibilities:
Take some time to explore the different post-processing options available in Eikon Diffusion. Experiment with different settings and see how they affect your images.
Eikon Diffusion implements DeepCache, a powerful feature designed to speed up image generation without sacrificing quality – as long as you use it wisely.
Think of DeepCache as a smart shortcut for AI. It recognizes repeating patterns in the image creation process and remembers them, so the AI doesn't have to waste time recalculating the same things over and over.
This clever trick, combined with leveraging the structure of the AI's "brain" (called a U-Net), allows DeepCache to significantly accelerate the time it takes to generate images.
Lets try:
- Check the "Use DeepCache" option in the DeepCache section of Eikon Diffusion's generation tab.
- Keep default parameters to first test:
Important Note: While DeepCache is incredibly effective, there are a couple of things to keep in mind:
- Lower "Cache interval" values generally lead to higher quality images, but also do not increase the speed of image generation.
- Lower "Branch ID" values may result in a slight decrease in image quality, while higher values may match the original image quality, so finding the right balance is critical.
| Original Image | DeepCache w/FreeU and without PAG | DeepCache w/FreeU and with PAG |
|---|---|---|
|
 |
 |
| 23.33 seconds | 12.75 seconds | 15.40 seconds |
You may need to readjust the other parameters already applied to get a better result. In our test, it was necessary to increase the PAG guidance scale from 2.0 to 3.0 to get a better result with PAG.
DeepCache is a great option if you need to generate images quickly and are willing to accept a small potential decrease in quality. However, if time is not a major concern, you can achieve the highest possible quality by disabling DeepCache.
Have you ever wanted to generate a series of images that share a consistent style, even when using different prompts? That's where Style Aligned comes in.
Imagine you're creating a set of images for a 3D game asset, and you want them all to have a consistent "macro photography" look.
How it Works
Style Aligned is like having a built-in style guide. Instead of providing a separate reference image, Eikon Diffusion uses the first image it generates as the style reference for all subsequent images.
- First Image: The initial image generated acts as the blueprint for the style.
- Subsequent Images: When you generate additional images, Style Aligned ensures that the new images maintain a similar "macro photography" aesthetic as the first generated image.
Benefits of Style Aligned
- Consistency: Create a set of images with a unified "macro photography" aesthetic, even when using different prompts describing various objects or scenes.
- Efficiency: No need to find a separate reference image – the first generated image automatically sets the style.
Let's explore the power of Style Aligned
- In the "Prompt" field, type one prompt per line:
a toy train. macro photo. 3d game asset
a toy airplane. macro photo. 3d game asset
a toy bicycle. macro photo. 3d game asset
a toy car. macro photo. 3d game asset
a toy boat. macro photo. 3d game asset.
- In Prompt method select "Normal".
- Select "DPM++ 2M Karras" from the Sampler dropdown menu.
- In the Seed" field, enter: 827980458.
- In Randomize Seed keep this unchecked if you want see this seed example or check it to random seed.
- Check Apply style aligned option.
Note: You can apply PAG if you want an improved result.
- Click the Generate button to generate images.
Style aligned without PAG
| train | airplane | bicycle | car | boat |
|---|---|---|---|---|
 |
 |
 |
 |
 |
Style aligned with PAG
| train | airplane | bicycle | car | boat |
|---|---|---|---|---|
 |
 |
 |
 |
 |
Style Aligned is a powerful tool that unlocks new levels of creative control in Eikon Diffusion. By ensuring consistency across a series of images, it empowers you to build cohesive visual narratives with ease.
Don't be afraid to experiment! Play with different prompts, styles to discover the full potential of Style Aligned. As you become more familiar with this feature, you'll find yourself creating truly remarkable and unified image sets.