SDXL Fooocus

About

Fooocus is an open source project built to make prompting SDXL models easier. It simplifies creating high-quality images by abstracting complex settings. It applies tricks under the hood to make "high quality" images with out needing to do extensive "prompt engineering". These tricks are applied under the hood to your prompt / images to generate the end results.

• Style System: Uses GPT-2 for prompt expansion, enhancing text representation and output variety.
• Refiner Swapping: Seamless parameter transfer within a sampler for image refinement.
• Negative Image Aspect Ratio Guidance: Compensates for lack of cross-attention in high-resolution outputs.
• Self-Attention Guidance: Mitigates smoothing effects in image outputs.
• LoRA Model Implementation: Enhances image quality with fine-tuned sampler parameters.

Features

• High Quality: Image generated are high quality due to the hidden tricks applied above.
• Easy to Use: No need to do extensive prompt engineering to get good results.
• NSFW: NSFW images are supported.

Limitations

• Hands, Feet, Faces, and Text: Sometimes miss the mark on hands, feet, faces, and text
• Prompt Understanding: Not as good as DALL·E 3 in following the prompt

Usage Tips

• Be Specific and Detailed: Include elements like setting, objects, colors, mood.
• Describe Mood and Atmosphere: Use words like "serene," "chaotic," "mystical."
• Use Descriptive Adjectives: For instance, "a fluffy, small, brown dog."
• Consider Perspective and Composition: Mention views like close-up, wide shot.