🧬 Klein's Anatomy: Male
We are on Matrix! Feel free to join to get progress updates, test/exclusive releases, and more.
Matrix: https://matrix.to/#/#the-weight-room:nope.chat
A highly detailed anatomical LoRA trained on natural language captions.
The current version is an alpha release, purely for testing and feedback.
📖 Model Overview
This model is not a standard "tag-based" LoRA. It leverages the advanced text encoder capabilities of FLUX to understand semantic context, lighting, and scene composition. The goal is to deliver superior anatomical accuracy - specifically for men.
Revamped Uncut and Cut
We dedicated nearly 1,000 hours across multiple training runs to develop this version of the model. We also developed custom tooling to create a cleaner, higher-quality dataset with more accurate captions.
Key Improvements
Strong emphasis on removing plastic-like skin textures.
Unified support for both cut and uncut penises in the same LoRA (uncut is experimental and may need more prompt guidance).
Trigger & Prompting Guide
Always use clinical, descriptive language for best results. However, tags work okay in this version.
Do NOT use slang like "cock", "dick", "shaft", etc. Stick to "penis" consistently.
For uncut: Include "uncut penis" in your prompt.
For cut: Include "circumcised penis" (or just "penis" if you want the model to decide).
For flaccid: Try "small flaccid uncut penis" or "small flaccid circumcised penis" to counteract any residual erect bias. This is still a work in progress.
Recommended Generation Settings
CFG Scale: 3.0 (keeps things natural and avoids overcooking)
Sampler: res_multistep
Steps: 25+
Aspect Ratio: Trained mostly on 9:16 (portrait/vertical) - works best for full-body or focused genital shots, but flexible horizontally too.
⚙️ Technical Specifications
Base Model: FLUX.2 Klein 9B-base (Use the fp8 version for best stability)
Trigger Phrase: circumcised penis (uncut penis for uncut) (!! READ MORE BELOW !!)
Training Res: 1024px (Native), samples generated at 1080p.
Use LoRA loader with clip connection for best results.
🎨 Prompting Strategy: "Natural Language"
🛑 STOP TAG SPAM.
Do not use comma-separated lists like: 1man, solo, realistic, bedroom.
✅ USE DESCRIPTIVE SENTENCES.
Prompt as if you are describing a photograph to a blind person. Focus on sentence structure, flow, and adjectives.
💡 The Semantic Anchor
While there is no "hard" trigger word, the phrase circumcised penis appears in almost every training caption.
How to use it: Integrate this phrase naturally into your sentence (e.g., "...showing detailed anatomy including a circumcised penis..."). This acts as a semantic anchor, signaling the model to prioritize anatomical stability and prevent "morphing."
📝 Sample Prompt
Copy this style for best results:
A photo of a 30-year-old man taking a shower behind clear glass. He is muscular with broad shoulders and defined abs, water running down his chest and glistening on his tan skin. He is washing his hair with one hand raised, while his other arm rests against the tiled wall. He is fully naked, and his flaccid circumcised penis hangs naturally. The lighting is bright and overhead, creating high contrast on his wet skin and muscle definition. He looks down with eyes closed, enjoying the warm water.
🚧 Development Status: Alpha/Beta
This LoRA is a Work In Progress (WIP).
Inconsistency: Some seeds are perfect; others may distort.
Lighting: Complex lighting can sometimes break anatomy. If you get errors, try simplifying to "natural light" or "soft studio lighting."
Updates: Expect behavior changes between versions as the dataset expands.
📸 Help Improve the Model
We need more data. The current dataset lacks selfie angles and handheld perspectives.
Requirements: You must be 18+, and must be the owner of the dataset (not randomly downloaded data).
How to Join: Join The Weight Room Discord server to contribute secure training data.
👉 Join The Weight Room Discord
⚠️ LICENSE WARNING: READ CAREFULLY
🛑 NON-COMMERCIAL USE ONLY
This LoRA is built on FLUX.2 Klein 9B, which is released under a strict Non-Commercial License.
❌ NO Paid Generations / Commissions
❌ NO Patreon / Subscribestar / Monetized Feeds
❌ NO Stock Photography Sales
By using this model, you agree to comply with the original FLUX.2 Klein 9B licensing terms.
Description
Experimental 4B lora model
FAQ
Comments (8)
this is really good, but it messes up eyes and kinda biases faces to like a certain type
I appreciate the heads-up. I'm currently training an expanded version to try and solve some of the issues that the current version has. To get better generalization without losing quality, I'm using a much higher rank, so the LoRA size is coming in at around 1.3GB. It looks promising, but to be sure, could you tell me what specific face types you are trying to prompt? I can check if the new version handles them better.
@oh_txt i think any prompt with the lora weight @ 1 reduces face variation and eye colors. usually to get better generalization/reduce concept bleed you reduce the lora rank and increase how many images youre training on 😩 but if its working i cant complain
@wiiz Are you in the Discord? I would love to get your feedback on the current training run
@oh_txt yeah sure dm me your dc
Does "Improved penis sizes" mean that "normal size is now possible too" or "even bigger"
Shouldn't be like 2 foot long now. You can also prompt small to help control
Thank you very much for your work. It's a very good Lora. Thank you.