Yuuka

Please like and share this model with friends! ♥️ I would love to see your creations, so please tag the model when you post them! If you enjoy this model, please consider supporting my work at https://lilougummy.gumroad.com/coffee or by tipping buzz! Thank you for your support!

Important Notes Before Use

• ⚠️ Epsilon Prediction (eps-pred): This model uses an eps-pred architecture.

• 🎨 Workflow Information: All sample images were generated using ComfyUI. The prompts and tags used can be found in the posted images to help you get started.

How to Use This Model: A Flexible Base

Please understand that Yuuka is a base model, not an aesthetically fine-tuned one. It is designed for power and flexibility, but this means it can be difficult to use with simple prompts.

• LoRAs and Artist Tags are Recommended: To achieve a specific style, it is highly recommended to use LoRAs or include artist names in your prompt.

• Avoid Basic Prompts: Simple prompts without stylistic guidance will likely not yield the best results. This model shines when given detailed instructions.

Model Motivation

This model, Yuuka, is an evolution of Illustrious 2.0. The primary goal was to build a model with a significantly stronger understanding of general NSFW concepts, character interactions, and complex poses. By refining the training data and captioning process, Yuuka aims to provide a more intuitive and powerful experience for generating detailed scenes.

Training and Dataset

Yuuka was trained on a meticulously curated dataset of 1.5 million images. The captioning process was designed to build layers of understanding:

• Base Danbooru Tags: Started with a solid foundation of standard tags extracted from Danbooru (ratings: questionable, explicit).

• PixAI Tagger Enhancement: The dataset was then processed with the PixAI tagger to capture any missing tags.

• Natural Language Captions: Descriptive, human-readable captions were added to improve prompt comprehension.

• Experimental Action Tagging: A custom #action -> #source -> #target captioning system was implemented to better define interactions between subjects. (Note: This feature is experimental and while it shows promise, it may not always function as expected. Further training is needed on this part.)

Key Improvements

• ✅ Enhanced NSFW Knowledge: The model has a much deeper and more reliable understanding of a wide range of NSFW concepts and scenarios.

• ✅ Improved Action & Poses: Yuuka excels at interpreting prompts related to actions between characters and generating more dynamic and accurate poses.

• ✅ Superior Prompt Comprehension: Thanks to the multi-layered captioning, the model is better at understanding and executing complex and natural language prompts.

• ✅ Based on Illustrious 2.0: Inherits the strong foundational aesthetic of Illustrious 2.0 while pushing its capabilities further.

Recommended Settings

• Steps: 30

• CFG Scale: 5

• Sampler: Euler

• Scheduler: Beta

• VAE: Baked-in

Feedback & Support

• 🐞 Report Issues: If you encounter any issues or have suggestions for improvement, please let me know! This is an actively developed model, and your feedback is invaluable.

• 🚀 Coming Soon: Keep an eye out for a full workflow guide and potential future versions with further refinements.