Wan2.2 i2v-A14B LoRa training has no effect. Comparison with other trainers AI-Toolkit

[wzgrx]

Device: Windows 11 4090 24GB 96GB RAM
Video material, resolution below 960, 16fps, duration 5s

Training 2D motion effects with the same material and settings, using AI Toolkit, LoRa works fine, but Musubi Tuner doesn't.Need help from experts and professionals to successfully implement musubi-tuner training lora and achieve the effect of using AI Toolkit. The following includes all the settings

video dataests
https://huggingface.co/datasets/wzgrx/2D-animation-effects

using AI Toolkit trian lora
https://civitai.com/models/1920897/wan22-2d-animation-effects-2d
Test results
https://github.com/user-attachments/assets/37332aa7-6286-49ef-9241-157bb82b7580

using musubi-tuner

same video

[general]
caption_extension = ".txt"
batch_size = 1
enable_bucket = false
[[datasets]]
image_directory = "E:/AI/musubi-tuner/train/image"
cache_directory = "E:/AI/musubi-tuner/train/cache_image"
resolution = [1024, 1024]
num_repeats = 1
[[datasets]]
video_directory = "E:/AI/musubi-tuner/train/video"
cache_directory = "E:/AI/musubi-tuner/train/cache_video"
frame_extraction = "uniform"
source_fps = 16.0
target_frames = [57]
max_frames = 57
enable_bucket = true
bucket_no_upscale = false
resolution = [384, 384]

python src/musubi_tuner/wan_cache_latents.py
--vae "E:\AI\musubi-tuner\wan\wan_2.1_vae.safetensors"
--batch_size 2 `
--i2v--dataset_config "E:\AI\musubi-tuner\train\2.toml"--vae_cache_cpu

python src/musubi_tuner/wan_cache_text_encoder_outputs.py
--t5 "E:\AI\musubi-tuner\wan\models_t5_umt5-xxl-enc-bf16.pth"
--batch_size 4--dataset_config "E:\AI\musubi-tuner\train\2.toml"--fp8_t5

accelerate launch --num_cpu_threads_per_process 1 src\musubi_tuner\wan_train_network.py
--dataset_config "E:\AI\musubi-tuner\train\2.toml"
--vae "E:\AI\musubi-tuner\wan\wan_2.1_vae.safetensors"
--flash_attn --split_attn
--fp8_base --fp8_scaled
--gradient_checkpointing
--network_module networks.lora_wan --network_dim 16 --network_alpha 16
--min_timestep 900 --max_timestep 1000
--max_train_epochs 20 --save_every_n_epochs 2 --seed 42
--log_with tensorboard
--blocks_to_swap 16 `
--offload_inactive_dit--task i2v-A14B--dit "E:\AI\musubi-tuner\wan\wan2.2_i2v_high_noise_14B_fp16.safetensors"--t5 "E:\AI\musubi-tuner\wan\models_t5_umt5-xxl-enc-bf16.pth"--mixed_precision fp16--optimizer_type adamw8bit --learning_rate 2e-4--max_data_loader_n_workers 8 --persistent_data_loader_workers--timestep_sampling shift --discrete_flow_shift 5.0--preserve_distribution_shape--output_dir "E:\AI\musubi-tuner\out" --output_name "my_i2v_lora_high"--logging_dir "E:\AI\musubi-tuner\train\log"

accelerate launch --num_cpu_threads_per_process 1 --mixed_precision fp16
--task i2v-A14B
--dit "E:\AI\musubi-tuner\wan\wan2.2_i2v_low_noise_14B_fp16.safetensors"
--vae "E:\AI\musubi-tuner\wan\wan_2.1_vae.safetensors"
--fp8_base --fp8_scaled
--gradient_checkpointing
--network_module networks.lora_wan --network_dim 16 --network_alpha 16
--timestep_sampling shift --discrete_flow_shift 5.0
--preserve_distribution_shape
--output_dir "E:\AI\musubi-tuner\out" --output_name "my_i2v_lora_low"
--log_with tensorboard
--offload_inactive_dit.\src\musubi_tuner\wan_train_network.py--dataset_config "E:\AI\musubi-tuner\train\2.toml"--xformers--t5 "E:\AI\musubi-tuner\wan\models_t5_umt5-xxl-enc-bf16.pth"--optimizer_type adamw8bit --learning_rate 2e-5--max_data_loader_n_workers 8 --persistent_data_loader_workers--network_args "loraplus_lr_ratio=4"--min_timestep 0 --max_timestep 900--max_train_epochs 20 --save_every_n_epochs 2 --seed 42--blocks_to_swap 18--logging_dir "E:\AI\musubi-tuner\train\log"

After a period of training, the generated lora is completely different from the lora generated using AI Toolkit. The lora trained with musubi-tuner has no effect at all. What should I modify? Thank you very much