LoRA trained on Wan2.1-I2V-14B-480P does not work

[nextl368]

Version: 1.1.8

I am encountering an issue where a LoRA trained on the Wan2.1-I2V-14B-480P (Image-to-Video) base model does not produce the expected video output.

Successful Scenario (Working)

• A LoRA was successfully trained using the Wan2.1-T2V-14B (Text-to-Video) base model.
• The video generated using the T2V-trained LoRA exhibits the desired effect.
• Crucially: This T2V-trained LoRA also works correctly when loaded onto the Wan2.1-I2V-14B-480P base model.

Problem Scenario (Not Working)

• A LoRA was trained directly using the Wan2.1-I2V-14B-480P base model.
• When this I2V-trained LoRA is applied, the generated video does not exhibit the desired effect.

Training and Data Details

• Training Commands: I used the exact same command line given in the WanVideo examples section for both the T2V and I2V training runs.
• Dataset: The dataset is identical for both training attempts. It consists of an input image and an associated prompt with just the trigger keyword.
• Metadata Example: I am deliberately omitting a specific descriptive prompt in the metadata.csv to focus on the style keyword:

video,prompt
training_image_1.png,"N0Y1V2R3S" 
training_image_2.png,"N0Y1V2R3S"
...
training_image_250.png,"N0Y1V2R3S"

ComfyUI Usage (Context, Not the Issue)

For testing, I used ComfyUI, ensuring I loaded the LoRA via the WanVideo-specific nodes. (Note: Standard ComfyUI LoRA loading nodes fail due to key name differences in the safetensor file produced by DiffSynth-Studio, but I have verified this is not the root cause of the I2V training failure.)

Summary of the Problem

The core issue is that I2V-native LoRA training is failing to capture the desired style/effect, even though the same dataset, training command, and environment allow for successful T2V LoRA training that can then be applied to the I2V model.

[AshkanTaghipour]

The same, not wrking for wan22-14B-Animate. get the bellow warning, i checked the grads and all lora grads were None! UserWarning: None of the inputs have requires_grad=True. Gradients will be None

[nextl368]

The same, not wrking for wan22-14B-Animate. get the bellow warning, i checked the grads and all lora grads were None! UserWarning: None of the inputs have requires_grad=True. Gradients will be None

I don’t think this is my case. I wrote a script to compare the LoRA from the T2V and I2V base models. Structurally, they are identical , both files contain the exact same set of keys. I also checked tensor types and dimensions, and there’s no issue there. Then I inspected the tensors themselves to detect any NaN or Inf values, all good. After that, I verified the LoRA weights for unusual patterns, such as zero standard deviation or abnormally high magnitude. Everything looks normal.

To ensure ComfyUI wasn’t causing the problem, I also ran inference using the commands provided in the LoRA Training Validation table. Still the same issue, the LoRA trained with the I2V produces no noticeable effect.