用新出的I2v lora Wan2.2-Lightning跑wan_moe_i2v_distill结果很奇怪

[Tinyyuumi]

这是分别用两个lora跑的结果： 1、wan2.2-lightning： https://github.com/user-attachments/assets/ff518b0f-6034-4108-abbd-64e248321db2

2、wan2.1 lora（高低噪声模型都加lora）： https://github.com/user-attachments/assets/71f5cce4-6821-4c72-b48a-92c0b72274c9

是框架不兼容吗，为啥一个结果那么差，大佬们有没有时间可以看看

config 配置如下：wan_moe_i2v_distill.json { "infer_steps": 4, "target_video_length": 97, "text_len": 512, "target_height": 832, "target_width": 480, "self_attn_1_type": "flash_attn3", "cross_attn_1_type": "flash_attn3", "cross_attn_2_type": "flash_attn3", "seed": 42, "sample_guide_scale": [1.0, 1.0], "sample_shift": 3.0, "enable_cfg": false, "cpu_offload": false, "offload_granularity": "model", "boundary_step_index": 2, "denoising_step_list": [1000.0000, 937.5000, 833.3333, 625.0000], "use_image_encoder": false, "lora_configs": [ { "name": "low_noise_model", "path": "Wan2.2-Lightning/Wan2.2-I2V-A14B-4steps-lora-rank64-Seko-V1/low_noise_model.safetensors", "strength": 1.0 }, { "name": "high_noise_model", "path": "Wan2.2-Lightning/Wan2.2-I2V-A14B-4steps-lora-rank64-Seko-V1/high_noise_model.safetensors", "strength": 1.5 } ] }