This is an experimental Rank 128 release trained specifically to achieve deep, high-contrast ink graphics.
This release is a failed experiment. The underlying image generation engine simply lacks the capacity to properly understand and handle the human body and a tattoo simultaneously. Regardless of the training rank used, the system fails to deliver a convincing, natural result, turning what should be a tattoo into a flat, detached graphic or generating severe skin corruption.
Dataset & Training Specifications:
Dataset Size: 164 selected images.
Training Resolution: 1024x2024 pixels.
Available Versions: 1500 steps and 2000 steps.
Crucial Note on Showcase Images:
All published preview images are completely raw, unedited, and unprocessed generations. No post-processing, no inpainting, and no external touch-ups have been applied. It is vital for this release that you see the pure, unfiltered output of the model exactly as it is generated.
Usage Recommendations:
Controlling Ink Density: Due to the aggressive nature of Rank 128, the model can deliver high ink saturation that might result in a flat "sticker" look. To make the ink blend naturally into the skin, explicitly include transparency markers in your prompt (e.g.,
70% transparencyorsemi-transparent ink).Step Selection: The 1500-step version provides a more flexible blend with the base model. If you run the 2000-step version for heavy blackwork, it is recommended to lower the LoRA weight to around
0.4 - 0.5to maintain optimal skin texture.
Conclusion: No matter how you tweak the settings or what rank you train on, the image engine cannot properly synthesize a normal human body together with an authentic tattoo. This LoRA is released purely as-is for logging and archival purposes of an unsuccessful training run.
That being said, perhaps someone will still find it useful in this state for their own testing or specific use cases.