Intro：
This LoRA is trained on openly sourced images from the public community, focusing on furry male characters (bara). It is designed to enhance image details and scene complexity while incorporating personal aesthetic refinements, offering strong generalization. The model is not intended to replicate specific artists' styles or techniques, and its impact on the base model's art style is minimal, with slight variations possible depending on prompts or base models.
Usage Recommendations:
No specific trigger word is required
It is recommended to use it in conjunction with the following prompts: male focus, furry, furry male, bara, muscular male
While trained as a v-pred model, it is also compatible with EPS MODELS.
The model has a small impact on the base model's art style, and results may vary slightly depending on the prompts or base model used.
V3.2 Golden gossamer：
What’s New in This Version?
Expanded Dataset: The dataset has been increased to over 9,000 images. I've also removed some specific art style images that were prone to causing overfitting.
Zero Terminal SNR: This version incorporates Zero Terminal Signal-to-Noise Ratio during training. In theory, this should provide better adaptability to images that are much brighter or darker than average.
Usage Recommendations & Key Changes
Weight: This LoRA can be strong-willed at full weight. Please use it at a weight of 0.95 or lower for best results.
Artist Strings: Compared to previous versions, the enhancing effect on artist strings decreases as you add more of them. In extreme cases with many artist strings, the effect might be diminished. However, it is very responsive to short and simple artist strings. For the best outcome, try to keep your artist strings as concise as possible.
Details and Lighting: Despite the changes, the LoRA continues to perform well in rendering details and lighting.
User Experience & Creative Potential
The experience of using this version is noticeably different from previous ones, both in its handling and the resulting style.
Its performance can be quite varied. In some cases, it shows a great deal of creativity, interpreting ideas in prompts to achieve effects that were difficult to get before. In other situations, its output is more standard.
I highly recommend generating several images with random seeds to explore its potential, especially if you find the base style isn't being overfitted and locked in. You might be surprised by the results you can achieve.
As always, thank you for your support, and I hope you enjoy experimenting with this new version.
V3.1 Ravine stream：
What's New?
Expanded Dataset: The training dataset has been expanded to over 8,000 images.
Improved Regularization: I've adjusted some training parameters to enhance regularization, making its general characteristics similar to version v2.8.
Better Style Compatibility: Compatibility with flatter, cel-shading-like art styles has been improved. However, to achieve specific art styles, it's best to use it alongside artist strings. For best results, ensure your base model is responsive to artist string prompts.
Usage and Compatibility
This LoRA should perform well on most base models without issues. However, due to the nature of large datasets, there's a small possibility of minor data overlap with the training sets of some base models.
Potential for Overfitting: If you use the LoRA at a high weight (like 1.0) and notice issues such as distorted background perspective or problems with hands/fingers, you may be experiencing overfitting.
Solution: If this happens, try lowering the LoRA weight or reducing the CFG scale. This usually resolves the issue.
Working with Artist Strings and Other LoRAs
Artist Strings: Using this LoRA with artist strings is a great way to conveniently adjust the art style. For this version, I have specifically enhanced the effects for the artist strings "boogbogex" and "luodejun" (please see the example images for reference). If you use a prompt that includes these two artists, you may need to rebalance your weights accordingly.
Combining with other LoRAs: You can achieve some unexpected and interesting results by combining this LoRA with other style LoRAs.
Heads-up: If your characters start showing features you didn't prompt for, such as purple skin or star-shaped tiger stripes, it's a sign that your LoRA combination is overfitting.
Solution: Adjust the weight balance between the different LoRAs to fix this.
Thanks for checking out this update, and I look forward to seeing what you create