Hello everyone! This is a body modification Lora for female characters, trained on the Z-Image-Base model using Concept Slider. The goal of this Lora is to give female characters a full bust and a tiny wasp waist. It's still experimental and has room for improvement, but overall, it achieves the intended effect. Below, I'll share the training details and how to use it.
💻I've already set up a complete workflow for you. Just click this ➡️ link to quickly try out the Lora's different effects online.
🎁Bonus: If you're signing up for the first time, you can get 1,000 free RH Coins by using my link and the invite code ➡️rh-v1182. Plus, you'll get another 100 RH Coins for logging in daily.
✍Lora Training Parameters:
Training Tool: AI-TOOLKIT
Training Set Image Count/Resolution: 36 images at 1024 pixels
Linear Rank: 24
Batch Size: 2
Learning Rate: 2e-4
Timestep Bias: High Noise
Training Resolution: 1024
Target Class: woman
Positive Prompt: A woman with massive heavy bust, an incredibly tiny wasp waist, fabric on the bust stretched extremely tight, wearing extremely skin-tight clothing, fabric on the chest area shows obvious horizontal stretching and wrinkling, a deep shadow beneath the chest.
Negative Prompt: A woman with completely flat chest, no cleavage, thick wide waist, loose-fitting clothes without stretch, flat lighting.
Anchor Class: slender woman's face, thin arms, thin legs, normal human anatomy.
🚀Testing and Usage Instructions:
1. Training Strategy: I didn't plan to use the Z-Image-Base model for the final image generation. I only need the Lora to generate the intended body silhouette, and then I use the Z-Image-Turbo model later to refine the details. You can see this in my workflow, where I use dual sampling to create the final image.
2. Lora Strength and Prompts: There are two ways to use this.
Method 1: Use a normal prompt + Lora (Strength: 1.5). This achieves the goal, but a strength of 1.5 might increase the chance of anatomical errors.
Method 2: Use positive descriptions in your prompt (e.g., "She has a full, round bust, prominent chest, obvious cleavage, fabric on the chest is noticeably stretched, tiny wasp waist") + Lora (Strength: 1.2). In this case, the Lora acts more like an amplifier, enhancing the "large bust + tiny waist" effect based on the model's original capabilities. This is the method I recommend. Either way, I suggest keeping the positive Lora strength between 1.2 to 1.5.
3. Negative Lora Strength: With a single-concept Lora trained using Concept Slider, a negative strength value should create the opposite effect. However, in my tests, negative values didn't produce a noticeable reverse effect. I suspect this might be related to my dataset images and the fact that the concept wasn't entirely "single." So, I don't recommend using negative strength values.
💖I'm E.K., an engineer who can't draw. I'm passionate about creating more fun and practical AI models.If you have any better ideas, let's discuss them!