A body proportion enhancement LoRA trained on FLUX.2 [klein] 9B Base and tested on the fast 4-step distilled model. It enhances curves and anatomy with a focus on realism without degrading clothing textures, poses, lighting, or overall prompt adherence.

Prompting & Behavior

• No Trigger Word Required: This LoRA reinforces standard natural language descriptors rather than relying on a custom token.

• Effective Descriptors: Works best with prompts like busty, very busty woman, hourglass figure, large breasts, etc. Experiment with standard descriptive phrasing to find what fits your scene.

• Dataset Scope: Trained entirely on clothed subjects across a wide range of outfits (casual wear, t-shirts, dresses, bikinis) to ensure fabrics drape and fold naturally. Due to the dataset distribution, there is a slight natural bias toward form-fitting clothes.

• LoRA Stacking: Stacks well with other LoRAs (adjust weights as needed). While this is not inherently an NSFW LoRA, it has been tested alongside SNOFS v1.4 with solid results.

Recommended Settings

• 4-Step Distilled (FLUX.2 [klein] 9B):

  • LoRA Weight: 0.7 – 1.0 (going over 1.1 may introduce anatomical distortions)

  • Steps: 4

  • Guidance / CFG: 1.0

  • Sampler: Euler

  • Scheduler: Simple

• Full Base Model (FLUX.2 [klein] 9B Base):

  • Trained on the full base model, but not extensively tested there. Lowering the LoRA weight is likely a good starting point.

Community Feedback & Testing

Primary testing was done using the 4-step distilled pipeline. If you experiment with the full 9B base model or find great LoRA stack combinations for either version, please share your settings, workflows, and sample generations in the gallery and comments!