an@t0my,
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 requiredIt is recommended to use it in conjunction with the following prompts: male focus, furry, furry male, bara, muscular maleWhile 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)CK5.5 DistantHowl：The overall characteristics and visual outputs of v5.5 are very similar to v5.2, even though they were baked using completely different training parameters. Please feel free to test both and choose the version that fits your workflow best.From a technical standpoint, this version was "low-temp baked" using an extremely low learning rate of 0.00000489. Why so low? Because the two attempted versions right before this one completely died during training. One was completely fried by the end of the first epoch, and the other finished fitting during the warmup phase and immediately started degrading. It turns out that the DoRA + LoHA combination is incredibly sensitive when dealing with large datasets.As for why I stubbornly stick to this specific architecture—it's primarily to make up for the accident that was version 4.5. Now, a friend refactores the "spaghetti code" of the trainer I was using (prior to that fix, any LyCORIS + DoRA combination I tried would always result in broken outputs). The limitations of the trainer at that time, combined with other unforeseen issues, made v4.5 notoriously difficult to use. This version is meant to redeem that architecture.Thank you all so much for your continuous support and patience!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 OverviewThis 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 UsingBecause 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: LoHAdtype: 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 EditionThis 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 WordsThe 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 GenerationThe 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 StyleOn 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.AcknowledgementsA 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. E
