Description

This is a test LoRA model trained on HunyuanVideo using my favorite character.

This LoRA depicts the Character Suou Amane from Grisaia series.

Trigger word: "suou_amane, red hair, long hair, blue eyes, "

Example: "suou_amane, red hair, long hair, blue eyes, an anime girl is looking at viewer. She is wearing school uniform, and smiling to viewer."

Usage Notes

Since this model was trained with a resolution of 960x544, it is recommended to generate videos in similar landscape aspect ratios for optimal results.

Training Info

Used kohya's musubi-tuner for training. https://github.com/kohya-ss/musubi-tuner

The main training parameters:

• Learning Rate: 1e-3

• Optimizer Type: adamw8bit

• Network Dim (Rank): ??

• Network Alpha: 1

• Timestep Sampling: sigmoid

• Discrete Flow Shift: 1.0

• Max Train Epochs: 10

• LR Scheduler: cosine_with_min_lr

• LR Scheduler Min LR Ratio: 0.1

• LR Warmup Steps: 0.01

Total training steps: 13550

Dataset Info

107 images and 230 video clips.

The frame extraction settings for videos in musubi-tuner are as follows:

frame_extraction = "uniform"
frame_sample = 4
target_frames = [1, 13, 25]