This is a low-noise LoRA and should be applied only in low-noise mode.
The recommended strength is 0.7, though 1.0 is also acceptable.
The showcase videos were generated using smoothmix t2v and remix t2v.
Training configuration
Realtime LoRA Trainer (Wan 2.2 – Musubi Tuner)
images: 30
captions: generated with Qwen3-VL-8B following a consistent rule set
noise_mode: Low Noise
training_steps: 600
learning_rate: 0.0003
lora_rank: 16
vram_mode: medium (1024px), FP8
This model was trained during testing of comfyUI-Realtime-Lora.
The settings listed above are the values I found to be optimal through experimentation.
If you are aware of configuration values that can produce better results, I would appreciate it if you could share them.