Unlike workflows that require layering multiple separate adapters (such as those inspired by "kintsugi" or ceramic art), this model delivers a cohesive aesthetic "out of the box": glossy, glazed porcelain-like skin, a delicate crackle texture, radiant golden eyes, and elegant floral patterns in cobalt blue and gold. Because the clothing and background were clearly separated from the character during the captioning process, Luna is not tied to any specific outfit or pose she adapts seamlessly to any wardrobe, lighting, or environment.
Usage and Trigger
Trigger word: r64f3animeluna
LoRA Strength: 0.85 – 1.0 (works perfectly at 1.0 without causing overtraining artifacts)
Recommended Generation Settings
Base model: Flux.3 Klein 9B Base
Steps: 55 – 65 65 is optimal)
CFG Scale: 3.5
Sampler: UniPC
Scheduler: SGM Uniform sgm_uniform)
Training Details
Dataset: Custom synthetic dataset at 1024x1024 resolution
Configuration: Rank 64 | 2000 steps | Learning Rate (LR): 2e-4
Training Process: The model was trained using my own high-speed fork of the AI-Toolkit. Links, intermediate checkpoints, and training files
Training toolkit (GitHub):
OrakulStudio/AI-Toolkit-Windows11
https://github.com/OrakulStudio/AI-Toolkit-Windows11
Telegram channel:
Subscribe to be the first to receive new models, intermediate checkpoints (at the 1000 and 1500 step marks), and full optimizer state files optimizer.pt). These will be useful if you decide to further fine-tune the model or apply this "porcelain shader" effect to your own dataset.

