This LoRA was trained at a low rank to minimize interference with different base models. Because it is a general-purpose LoRA, the output can vary depending on the base model, number of steps, CFG (guidance scale), sampler, and scheduler.
It would be very helpful if you could read all the information below

☕ Support

If you find this LoRA useful, your support is always appreciated:
https://buymeacoffee.com/lacoz

📌 Training Details

- Base Model: Stable Diffusion 1.5
- LoRA Type: Face LoRA (UNet only)
- Resolution: 512×512
- Epochs / Steps: 30 epochs / 8640 steps
- Batch Size: 2
- Repeats: 2

⚙️ LoRA Settings

- Dim / Alpha: 128 / 40
- Learning Rate (UNet): 3e-4
- Optimizer: AdamW 8-bit
- Scheduler: Cosine + 7% warmup
- Precision: FP16
- min_snr_gamma: 5
- noise_offset: 0.02

🔹 Test

I tested the following models :)

AsianVibes_NSFW,
HirosakiMix＿NSFW,
JunX Asian Dreamix,
real asian beauty,
Sultry Grace Asian,
WildCardX -Asian,
WildCardX REAL NSFW SFW

Try other models as well :)

🧪 This is how I use it

- Sampler: DPM++ 2M
- Scheduler: Karras
- Steps: 30
- CFG Scale: 3 ~ 7
- LoRA Weight: 0.65 ~ 0.9
- Latent Resolution: 512×768 (or 768×512)
Note: Optimal values may vary depending on the base model, so feel free to adjust these settings to suit your setup.

🔹 Trigger

Feel free to use the trigger—or not—whichever works best for your workflow.

When using the lacoz trigger, the LoRA’s identity becomes stronger and more defined.
Even without the trigger, the face blends naturally into the base model and remains consistent.

🔹 Lora Weight
The recommended LoRA weight is fixed at 0.7~0.9

f the LoRA face seems to have little or no effect with your base model, recommend increasing the LoRA weight gradually.