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.
Due to the training on specific character features, you might occasionally see a particular pattern appear.
- To avoid diamond-shaped body markings/tattoos, please add the following to your negative prompt---especially the tiger furry: diamond (shape)

CK5.2 FallBlossom：

1. Return to Ease of Use You can think of this version as a much more user-friendly alternative to v4.5. I've completely removed the complicated usage conditions, bringing it back to the plug-and-play simplicity of earlier versions. Best of all: no trigger words are required.

2. Subtle but Visible Enhancements On most standard base models, this LoRA provides a noticeable but unobtrusive improvement to the overall image quality. These enhancements can apply to various aspects of the generation. As a reminder, the original purpose of training this LoRA was to compensate for the regression in furry concepts found in the chenkinnoob05 model.

3. Unexpected Synergies When paired with certain unique or highly stylized base models, you might notice some bizarre or unexpected "chemical reactions." If you encounter wild or highly unique generations under these conditions, don't worry—this is completely normal and part of how the model interacts.

4. Massive Dataset & Tagging Overhaul The training dataset has now surpassed 30,000 images.

About 1/10 of these (3,000 carefully filtered down from an 80,000-image pool) are high-quality furry male images sourced from Danbooru using standard tags.
For the rest of the dataset, I appended e621 tags alongside specialized, abstract concept tags generated by Gemini during the captioning process. This experimental tagging approach is explicitly designed to improve the LoRA's responsiveness to the highly imaginative, descriptive, or "hallucinated" prompts generated by LLMs.

V4.5 Overtuned：

Model Overview

This is a specialized DoRA (Weight-Decomposed Low-Rank Adaptation) model, built using the LoHA architecture and trained with V-Prediction.

My testing confirms that it subtly influences the style and details without destroying the original composition or structure of the base model. It is ideal for users who need precision rather than drastic changes.

⚠️ IMPORTANT: Read Before Using

Because this is a V-Pred LoHA-DoRA, it behaves differently from standard SDXL LoRAs. It requires specific loading methods to work correctly.

Current Compatibility Status:

❌ SD WebUI Forge: Currently does not work/not supported.
✅ ComfyUI: Fully supported (see Method 1).
⚠️ Standard Loaders: Supported with conditions (see Method 2).

🛠️ How to Use (Usage Guide)

There are two known methods to use this model effectively. Method 1 is recommended for the best quality and compatibility across different base models.

Method 1: The "Full Experience" (ComfyUI Only)

Recommended for: Best quality, usage on both Epsilon & V-Pred checkpoints.

To use the full capabilities of the DoRA architecture (Magnitude + Direction), you must use the custom LyCORIS Loader node.

Extension Required: ComfyUI-LyCORIS by AbstractEyes

⚙️ Node Settings:

strength_model: 0.01 to 0.1 (See "Weighting Note" below)
strength_clip: 0.0 (TE is not needed for this method)
lycoris_type: LoHA
dtype: bfloat16

📝 Critical Usage Notes for Method 1:

Weighting is extremely sensitive: The "Sweet Spot" is below 0.1. Start at 0.0 and increase by 0.01 increments until you reach the desired effect. Do not use 1.0.
Chaining Order: Due to DoRA characteristics, this node must be placed BEFORE any standard LoRA Loader nodes in your workflow.
Performance: You may notice slower load times during the first generation or when switching checkpoints/weights. This is normal behavior for the LyCORIS loader and does not affect the final image quality.

Method 2: The "Standard" Method (V-Pred Checkpoints Only)

Recommended for: Convenience, or if you cannot use custom nodes.

You can load this model using the standard LoRA loader (or LoRA Stack), but ONLY if you are using a V-Prediction Base Model.

Weighting: Functions like a standard LoRA. Recommended range: 0 - 0.8.
⚠️ WARNING: Do NOT use this method with Epsilon-Prediction models (standard SDXL, Juggernaut, etc.). It will result in severe overexposure, artifacts, and noise.

V3.9 BioGenesis：

The basic usage remains the same as in version 3.6, with just one new trigger word added: "s4mm2r". This was trained on images related to summer themes—yeah, there's an overwhelming amount of furry content in that space, so I decided to train it as a standalone concept.

I've switched from LoKR to LoHA, expanded the training set to over 20,000 images, and made it compatible with generating adult male humans. This should help improve the ability of certain Illustrious models (which might lack strong furry concepts) to produce bara-style male furries.

Compared to previous versions, this one has a lower degree of fitting and doesn't have a pronounced unique style. When used alone, it can generate more diverse images across different seeds, as long as your base model isn't locked into a very rigid style. That means it shares the easy-to-tune qualities of version 2.8, and it's even more responsive and flexible. However, it loses the representation of specific characters, so you won't have to deal with those diamond-shaped body patterns anymore.

Note: This version differs in the matching degree of artist strings when using the Vpred version base model compared to the eps version base model. Please experiment and adjust accordingly based on your model.

Before getting to this, version 3.8 went through a lot of failed attempts. I spent quite a bit of time digging into academic papers on training to figure things out. Now, version 3.9 might not stand out as much or feel as unique compared to earlier ones—there's always a trade-off between diversity and fidelity. I'd love to hear your feedback, and hope you have fun with it!

ill4.0 Sattva：

Illustrious Special Edition

This is a lightweight version optimized for Illustrious.

Note: The overall effect is generally inferior to the "Noob" version. Only switch to this version if the "Noob" version performs poorly with your specific Illustrious-based checkpoint.
Behavior: Similar characteristics to v3.9, with a slightly different generation feel.

V3.7 Vishvarupa：

Usage and Trigger Words

The basic usage is the same as version 3.6.
It utilizes the same nine trigger words as before, which you can freely mix and match.

Multi-Person Image Generation

The training images containing multiple people were specifically categorized and processed for this update.

Improved Stability: After multiple rounds of testing, I've found that when this LoRA is used on its own, it does not tend to cause severe issues like broken fingers or multiple heads, even when pushed to a weight that overfits its style.
Troubleshooting: If you are still encountering these kinds of structural problems, the cause is likely one of the following:
- Your Base Model: Ensure the base model you're using has good anatomical capabilities on its own.
- LoRA Combination: The issue might be caused by overfitting from other LoRAs you are using alongside this one.
- Your Tags: Review and adjust your prompts, as your tags can also influence the final structure.

Weighting and Style

On some base models, the style is quite strong at a weight of 1.0.
At a weight of 0.8 or below, the style is more subtle, making it ideal for everyday use or for blending with other LoRAs without overpowering them.
The model continues to pair very effectively with artist strings.

Acknowledgements

A special thanks to Soxan for their support and assistance. They provided valuable inspiration that helped improve the model and also assisted with testing.

V3.6 Cumulonimbus：

This new version features an optimized training set, resulting in a different base style compared to previous releases. I've included 2.5D style images in the training data, which has significantly improved the LoRA's material generalization. You'll find that textures for objects, skin, etc., will now adhere more closely to your base model's inherent style.

The biggest change in this version is the introduction of 9 distinct trigger words. Each one is trained on a different concept and carries a slightly unique style, texture, and effect. You are encouraged to mix, match, and adjust their weights to find the perfect blend for your creations.Here is a breakdown of each trigger word, based on the training data and my own experiments:

* an@t0my: Trained on images with excellent anatomical detail. This is intended to help improve and correct anatomical errors in your generations.

* bu2nF1ght: Trained on images with a red-heavy color scheme or combat themes. Recommended when creating images with these subjects.

* chr0m@t1cAbe2r@t1on: Trained on images featuring chromatic aberration or pale/pastel color palettes. Recommended for achieving these specific visual effects or moods.

* c0nc8pt: Trained on images with strong character designs, including cowboy shots, busts, and full-body portraits. Use this to create well-defined character concepts.

* d@rkF@nt@sy: Trained on dark fantasy-themed images. Recommended for that genre.

* F@nt@sy: Trained on images with a distinct fantasy or hand-drawn/painterly scene style. Great for creating fantasy environments.

m18ed,: Trained on a diverse mix of themes, with a significant portion being suggestive or *NSFW**. Styles range from flat color to 2.5D.

* ⚠️ Important Note: A large number of images featuring the character 'Sadiend' were included in this set, so this trigger may cause some overfitting towards his appearance. Use with caution if you do not want to generate images of this character.

* P@le2e: Trained on images with a strong concept art aesthetic. Recommended for creating concept art-style pieces.

* ske2chL1ne@rt: Trained on line art, sketch-style, or images with strong, defined lines. Recommended for achieving a line art or sketch-like effect.

You can get a general idea of the styles from the comparison grids below, but I highly recommend experimenting to see how they work with your own prompts and models.

[![<sup>1</sup>](https://i.postimg.cc/qBsVhw5x/xyz-grid-0013-630922844.png)](https://postimg.cc/3WxbzX7k)

[![<sup>2</sup>](https://i.postimg.cc/0QCnxzVD/xyz-grid-0015-974726991.png)](https://postimg.cc/R6NKLZ5Z)

Hope you enjoy the update, and happy generating.

V3.5 Twilight：

What's New?

Restored Functionality: Some features from older versions have been reinstated for improved usability. This has unavoidably increased the file size, but hopefully, it's still manageable(doge).
Greater Base Model Influence: The final image's foundational style and the impact of artist strings are now more heavily influenced by the base model you use. This gives you more control if your base model is well-trained.
Broader Optimizations: Compared to the last release, this version offers more noticeable optimizations in several areas, including:
- Art style consistency
- Body structure and anatomy
- Character facial features

Important Usage Tip

Due to the training on specific character features, you might occasionally see a particular pattern appear.

To avoid diamond-shaped body markings/tattoos, please add the following to your negative prompt: diamond (shape)

Thanks for trying it out, and I hope these changes improve your creative workflow.

V3.3 Glittering gravel：

Key Technical Changes

Dataset Expanded: The dataset has been increased to over 9,500 images.
New Training Method (LoKR): I've switched from LoCON to LoKR for this version. The most immediate benefit is a much smaller file size.
Noise Offset: This version introduces noise offset, which should provide a noticeable improvement in image contrast.
Dataset Refinements:
- To help reduce the smooth, plastic-like "AI feel," I've added a selection of real-world photography (without portraits) and removed images generated with the Illustrious model.

Performance and Usage

Quality and File Size: Don't worry about the smaller file size. In most cases, its performance in rendering lighting, details, and backgrounds is on par with previous versions.
Not a Style-Specific LoRA: Due to the training method and data variety, this LoRA does not strongly fit one particular art style or replicate specific characters perfectly. The advantage here is that it won't hijack your intended art style, giving you more creative freedom as long as you avoid overfitting.
Artist Strings: I have not observed any degradation when using artist strings. In fact, the overall results are often more flexible and diverse than in previous versions. This is likely because the LoKR training method reduces the model's rigidity, making it more versatile.

This version still excels at enhancing details and lighting, and now offers greater flexibility in style combinations.

Thank you for your continued support, and I hope you enjoy experimenting with this new version.

V2.8 Deneb：

This new version of the LoRA is focused on addressing some of the overfitting issues from the previous release and aims to have a more reasonable and improved impact on scene perspective.In most cases, the stylistic changes it introduces will be subtle adjustments to your chosen base model.

The more complex your prompts are, the more noticeable its effects will be. You should see clear improvements in composition, image layering, and the richness of the background.Conversely, with simpler prompts, the generated images will hew more closely to the output of the base model you're using.

What does this mean for you? This version is a bit more conservative in terms of raw creativity, but the trade-off is enhanced stability and predictability – it's more "obedient" to your intentions. My main goal in creating this was to make something that's genuinely "easy to use." You should find it significantly easier to correct errors and refine your images by tweaking your prompts.

Looking forward to seeing what you create with it!

Key improvements:

Reduced overfitting compared to the previous version.
More logical and improved scene perspective.
Subtle style adjustments that complement your base model.

How it behaves:

Complex Prompts: Expect noticeable enhancements in composition, depth, and background detail.
Simple Prompts: Results will be closer to your base model's standard output.

Overall:

More creatively conservative but significantly more stable and responsive.
Designed to be "easy to use" – simplifying the process of prompt-based image correction.

V2.0：

Dataset Update! 2000+ New Images Added

More than 2,000 new images have been added to the dataset, significantly boosting creative potential! This update also shows great results when mixed with other LoRAs.

Important Note: Some source materials were generated using Illustrious-series Checkpoints. If you notice an 'oily skin' effect in your generations, you can easily resolve this by adding the following to your Negative Prompt:

`(smooth skin:1.1), greasy skin, artificial gloss, plastic texture, nin_fake`

Happy creating!

Terms of Use for Model

Thank you for your interest in using this model! To ensure a positive and fair experience for everyone in the community, I’ve outlined the following terms. These are designed to protect both you as a user and me as the creator, while encouraging creative exploration. Please take a moment to read through them before using the model.

1. Usage Restrictions
This model is provided for creative and personal use. However, it must not be used for purposes that violate laws, ethical standards, or community guidelines. Specifically, the model is prohibited from being used for:

Creating or distributing content that is illegal, harmful, or offensive.
Any form of exploitation, harassment, or discrimination.
Commercial purposes without prior written permission from me.
Note: These restrictions are in place to maintain a safe and respectful environment for all users.

2. Copyright and Licensing
I retain full copyright ownership of this model. You are granted a non-exclusive, non-commercial license to use the model for personal or artistic purposes. Commercial use or redistribution of the model itself is not permitted without my explicit consent.

There are no restrictions on open-source merged models being used for image generation services, but it is recommended to credit the sources of any merged models.

3. Limitation of Liability
I am not responsible for any damages, losses, or issues arising from the use of this model. This includes, but is not limited to, unintended outputs, data loss, or compatibility issues with your hardware or software. I encourage you to exercise your own judgment when using the model and to ensure it meets your specific needs.

4. Use at Your Own Risk
By using this model, you acknowledge that you do so at your own discretion and risk. I cannot guarantee the model’s performance, outputs, or suitability for any particular purpose. Please ensure you understand the potential risks and use the model responsibly.

By using this model, you agree to abide by these terms. If you have any questions or need clarification, feel free to reach out to me through Civitai’s messaging system. I’m excited to see what you create, and I hope this model inspires your work!

Files with SHA256: 3fc0617002af5d458bd36a2744dd960cd1be3414b1ce876d2f165549ec381d8b