User Interface Guide

Welcome to the updated user interface. This guide is designed to provide you with a clear understanding of the recent UI changes and features. It aims to equip you with the knowledge you need to effectively utilize the updated UI. The UI includes mouse-over information for quick descriptions of each UI element.

UI Elements

Positive Prompt

The Positive Prompt window allows you to specify what you want to include in the image using various weights and special operators. (More information on this in Prompting 101).

Negative Prompt

The Negative Prompt window instructs the model on what not to include in the final image. You can access this feature in the advanced settings.

Image Generation

• Pro users have the privilege to create up to 4 images, while our valued free users are limited to 2.
• Images created with a seed and subseed are limited to 1 image.
• Images created with a set seed and subseed set to 0 have a max of 4.

Save To Folder

• You can now create folders directly from the image generation page.
• Images saved in folders will no longer be added to the general library.
• We recommend creating a new folder for each prompt or per day. Folders load faster than the library.

Width and Height

When setting the width and height for image creation, consider the following:

• Use portrait-like aspect ratios (512x765) for images requiring more vertical space
• Use landscape-oriented ratios (765x512) for wider, horizontal layouts.
• The choice of aspect ratio can dramatically impact the composition and visual appeal of your generated images.

Steps

You can now control the number of steps in image creation. With the new update, you can use up to 39 steps, which is a significant improvement from the previous limit of 20. More steps mean better blending, more terms, and clearer images. The minimum number of steps available is 10.

Seed

Seeds serve as the foundational elements in the image generation process. They help create the initial pixel patterns that eventually transform into images. Each time you specify a seed, you generate a unique matrix of random pixels. Setting a specific seed ensures that this random matrix remains consistent.

Subseed

Subseed is a secondary random number generator (RNG) that influences various aspects of the image. It maintains the same image structure as the main seed but introduces changes in elements such as coloring, placement of magic effects, hair, and other image qualities. Subseed play a crucial role in fine-tuning the image generation process. Subseeds enable you to explore subtle changes in specific aspects of the image. Whether you want to abjust characteristics of your image or test the effects of keywords, subseeds provide the means to do so effectively.

Subseed Strength

Subseed Power controls how much the original image differs from its variation.

UI Configuration Setups

Refining the Prompt (Fixed Seed)

Challenging prompts often require fine-tuning to achieve optimal results. This can be accomplished by employing a Fixed Seed. You have the option to generate a random seed by clicking the dice icon next to it or copying and pasting a value from a previous image. As a result, the model consistently generates the same image with each iteration. Customizing the Prompts, Steps, CFG Scale, and sampler settings introduces variation and aids in refining problematic prompts.

Variations of the Image (Subseeds and Subseed Strength)

Using a Fixed Seed locks the model into generating only a single image. Configuring the Subseed creates of the initial image. Please note that this setup is limited to producing a single image output. The Seed Strength can then be used to fine-tune the effect of the Subseed. Setting the Seed will consistently produce the same image across all four slots. However, when you increase the Subseed strength, it will introduce variations, resulting in four distinct images based on the same Seed.

Samplers

To best understand how the sampler works, you should first understand the specifics of how an AI image generator creates images. First, the image generator uses a random or specified seed to generate an array of latent noise that serves as the base of the image. Without any processing, this noise appears as something similar to the static you would see on television. The generator subsequently modifies the latent noise according to the positive prompt and negative prompt, emphasizing concepts specified in the positive prompt and removing concepts specified in the negative prompt. The sampler subsequently refines the modified noise over a certain number of steps to create the final image. Both the number of steps and the sampler used will therefore affect the appearance of the result.

Euler

Euler is the fastest and straightforward of the samplers. It was the original setting used on Sexy AI.

Euler A

Euler A akin to Euler but handles the latent noise differently.

DPM++ 2M Karras

DPM ++ 2M Karras handles longer prompt more effectively and is much more accurate than both the Euler/Euler A .

Sampler Comparison

Below you can see a comparison of two of the samplers available on sexy.ai and a demonstration of how samplers refine an image over a number of steps.

Euler	DPM++ 2M

CFG Scale (Classifier-Free Guidance)

The CFG scale value determines how closely the model follows your instructions. Lower values result in the model not following your instructions closely, while a high value ensures strict adherence to your prompt. To create a balanced image that aligns well with the prompt, start with a CFG value between 7 and 9. For more complex prompts, consider using a CFG scale of 12 to achieve higher alignment with the text.

CFG 1	CFG 7	CFG 12

FaceFix

FaceFix, an Automatic1111 GUI feature, designed to enhance facial features. It works effectively with photorealistic images but has limitations when applied to artistic style images. The following images used the same prompt, CFG scale, steps and sampler. Only Facefix was changed.

FaceFix (Off)	Facefix (On)

UI Refinement Process

1. Extract the Original Seed and Subseed Open the image and click on the information icon (i) in the top left corner. If your image contains them, you'll find the original seed and subseed displayed at the bottom of the menu.

2. Fix the Original Seed and Subseed This ensures the final image closely matches the original, maintains the same composition and details. This step is crucial for consistency.

3. If No Subseed Is Available You can create one by running the original seed through the model until it generates an image close to the original. Then, locate the subseed in the menu and fix it alongside the original seed.

4. Experiment with Different Models Engage with various models to explore diverse possibilities for your image.

5. Adjust Advanced UI Settings

6. Begin modifying the Advanced UI settings to craft a superior image.

Image Comparison

Positive Prompt

sketch pad, 3rd person perspective, from above, from behind, a mysterious sorceress, black background, illuminated by firelight, hairstick bun hairstyle, (best image:1.5), (best quality:1.5), (incredibly detailed:1.5), in a gothic style catacomb, cinematic volumetric backlighting, (extremely realistic:1.5), hell spawn enemies, action pose, a glowing ice spell striking an enemy, blue split complementary color palette, alcohol marker

Negative Prompt

3d, render, (worst quality:1.5), (worst detail:1.5)

Settings Changes

• Model Name: DreamShaper
• Image Size: Changed from 512x600 to 768x768 to create the closest camera angle to the ... original.
• Steps: Increased steps from 20 to 39 to add as much detail as possible.
• Seed: No change.
• Subseed: Original did not have a Subseed. Created a subseed using the method explained in UI Refinement Process
• Subseed Strength: This value was changed from 0 to 10, creating the back view and giving it the most detail.
• Sampler: Changed from Euler to DPM++ 2M Karras, creating the most balanced image that aligns with the original concept.
• CFGscale: Changed from 7 to 8; this was the best value. Values 1-6 created images with a loss of color, while 9 to 12 created oversaturated images.