Shuntea/yamashita_mitsuki_wan22_LOW

Yamashita Mitsuki - Wan2.2 T2V LoRA (Low-Noise)

Low-Noise Character LoRA for Wan 2.2 Text-to-Video model.

Available Checkpoints

This repository contains 10 checkpoints (Epoch 3, 6, 9, 12, 15, 18, 21, 24, 27, 30).

Recommended: Use yamashita_mitsuki_LOW.safetensors (Final version, Epoch 30)

Model Details

• Type: Low-Noise LoRA (Timestep 0-875)
• Base Model: Wan 2.2 T2V 14B (Low-Noise)
• Training Dataset: 30 images
• Training Epochs: 30 (7,200 steps)
• Network Dim/Alpha: 32/32
• Learning Rate: 1e-4
• Resolution: 768x768

Usage

Important: This LoRA must be used together with High-Noise LoRA

Required Files:

• High-Noise LoRA: https://huggingface.co/Shuntea/yamashita_mitsuki_wan22_HIGH
• Low-Noise LoRA (this file)
• Base Models: wan2.2_t2v_high_noise_14B_fp16.safetensors and wan2.2_t2v_low_noise_14B_fp16.safetensors

Settings:

• LoRA Weight: 0.8-1.0
• Use both HIGH and LOW together
• Compatible with ComfyUI

Training Details

• Optimizer: AdamW8bit
• LR Scheduler: cosine_with_min_lr
• FP8: Enabled
• Discrete Flow Shift: 5.0
• Training Time: ~5 hours on A100 80GB

Pair with High-Noise LoRA

For complete video generation, download and use both:

• High-Noise LoRA: https://huggingface.co/Shuntea/yamashita_mitsuki_wan22_HIGH
• This Low-Noise LoRA

Citation

Trained with musubi-tuner by kohya-ss.