Natalya is a virtual character created using a dataset generated by Gemini 3 Pro.

Natalya is a Russian woman with classic Slavic look with piercing deep blue "siren eyes" and a distinct small dark beauty mark located precisely above the left side of her upper lip.

Dataset of 50+ high-res images.

Training parameters:

Total Epochs = 12 (I started getting good sample results as soon as Ep 03)

Total Number of Steps = 3168

Text Encoder Learning Rate = 0.00001

Unet/DiT Learning Rate = 0.0001

LR Scheduler = cosine

Optimizer = AdamW8bit

Network Dim = 64 (needed more details)

Network Alpha = 32

Gradient Accumulation Steps = 1

Noise offset = 1

Multires noise discount = 0.3

Multires noise iterations = 6

Left rest of the values as is.

Regarding training steps, you don't need to train as long as I did. The loss rate plateaued by epoch 5 and stayed at the same level throughout.

Loss rate started with (0.284), dropped to 0.269 by Ep 3, then at Ep 5 it was 0.256; It stayed around this value until the end, at Ep 12 with 0.25.

This should mean that I could have stopped training anytime after Ep 05, and the results should have been fairly good.