SDXL User guide

Overview

SDXL is new version base model of stable diffusion, providing images of the highest quality at a fixed image resolution of 1024x1024. This guide aims to offer comprehensive information and recommendations for using SDXL (SFW), highlighting its key features and optimal usage.

This Model Excels In

• Generating SFW images.
• Creating a wide variety of high quality images.
• Efficiently handling simple prompts without the need for extensive positive/negative prompts.
• Generating outstanding images with clean details and great clarity.
• Recognizing art/artist styles.

What To Be Aware of

• The engine is slightly biased towards paintings and other artwork styles.
• Initial colors may appear dull, especially in photography, due to a stronger grayscale map for lifelike shadows.
• Not good at creating NSFW content.
• Model does not achieve perfect photorealism.
• Model does not support seeds/variations.
• The model struggles with more difficult tasks which involve compositionality, such as rendering an image corresponding to “A red cube on top of a blue sphere”.
• Faces and people in general may not be generated properly. This model was not trained to be factual or true representations of people or events.

Prompting Style

SDXL model works well with all stable diffusion prompting methods. This model has wide-range of term knowledge but is better at artistic usage.

Recommended Settings

• Steps: 28-39
• CGF Scale: 8
• Sampler: Euler (DPM++ 2M Karras)

Prompt Component Order

1. Art/Artist Style
2. Angle/view/style of image
3. Subject
4. Detail/quality terms
5. Lighting
6. Shadows
7. Additional modifiers

Example of Recommended Prompt Structure

closeup shot, portrait of a woman standing in the forest, ultra realistic, (best quality:1.5), extremely detailed, soft light, soft backlighting, soft shadows, (best image:1.5)

Sample Prompts

Stained Glass Desert

Positive Prompt : Stained glass art, a intricate landscape and camel, ultra detailed, crimson sunset, dawn, endless sky with sharp horizon, desert with several plants, light reflection, realistic perspective, depth of field

Negative Prompt : (worst quality:1.5), (worst image:1.5), (low detail:1.5), (low quality, bad quality), (blurry:1.2), grayscale, monochrome

SDXL Mecha Angel

Positive Prompt : (high quality, high detail:1.2), vray, (mecha angel), winged, glowing, (chiaroscuro:1.3), galaxy background,(glass), deep focus, (bokeh effect:1.2), soft ambient lighting, dramatic shadows, diffuse backlighting, (film grain:1.2), (best quality:1.5), (high contrast), floating particles,( antialiasing), cinematic shot, [red|orange|pink|deep blue] color scheme

Negative Prompt : figure, (repetitive patterns:1.5), low quality, normal quality, (worst quality:1.5), (worst render:1.5)

Withered Love

Positive Prompt : illustration, upperbody portrait, 1woman, abstract, surreal, (heartbroken:1.2), withered rose, dark ambient, highres, ultra quality, amazing background, (brush strokes), glowing rose petals, black theme, magenta theme, mist cover, gloomy, {by charlie bowater|by stephen gammel|by hans hartung|by simon prades}, sharp focus, kodak professional metallic, chiaroscuro, diffused lighting

Negative Prompt : (worst quality:1.5), (worst image:1.5), (low detail:1.5), (low quality, bad quality), lowres, red eyes, (blurry:1.2), body horror, ugly, wrong orientation

Prompt Style Comparison

Below you can see the difference in generation quality between applying only the general prompt style and applying the recommended style specific to this model.

Final Words

SDXL is a SFW model capable of a wide array of image styles, but it's worth noting that SDXL struggles with accurately rendering hands and may delve into excessive detail, which users will have to manage using the negative prompt. Additionally, challenges with skin texture may arise due to the SDXL refiner, which autocorrects the final image. Despite these drawbacks, SDXL increased resolution capability, and ease of multilpe character creation sets it apart.