⚠️ The v1 version has serious issues and is not recommended for use.

Core focus: a foot inserted and moving inside the mouth, and inserted

Version: v2 (based on v1)

Updates:

• Uses the latest dataset, improving both quality and stability.

• Fixes flickering, flashing, and green-screen issues.

Notes:

• Both high-noise and low-noise LoRAs were trained for 3000 steps. The high-noise LoRA performed best at the 2000-step checkpoint. Since the low-noise version showed no noticeable difference when paired with the high-noise LoRA across step counts, it is also released at the 2000-step checkpoint.

• For optimal results, it is recommended to use this model with other LoRAs and select a square or landscape resolution for the first frame.

Known Issue:

When using the sample video workflow, even if the input image contains fully covered without visible details, the generated video may still introduce small red dot–like artifacts that follow the motion and may occasionally also appear in the heel area.

As shown in the image below:

This issue is quite interesting and typically occurs only under very large motion intensity. It does not appear when using this LoRA alone.

Aside from this LoRA, I often use the same checkpoints and the "" LoRA from the sample video workflow for other motions, and the issue rarely occurs in those cases—so rarely that it can basically be ignored. Fortunately, I had previously saved a video where this issue did occur, which is shown below.

After loading this LoRA, the likelihood of this issue increases significantly. I initially suspected it might be related to the training dataset; however, videos with visible only account for around 10% of the dataset. Therefore, This issue is more likely due to a mismatch among this LoRA, the checkpoints, and the additional "" LoRA.

Recommended Solutions:

• Avoid loading additional LoRAs (e.g., "" LoRAs) in the sample workflow and use this LoRA independently.

• If a LoRA is required, consider upgrading the checkpoints to V10 or using alternative checkpoints.

Both approaches will reduce overall motion intensity. The issue has mostly been resolved.

if you want to preserve strong motion, you may need to accept that small -area details (nips) will appear in the initial frame.

Major issues in v1.

The output is sensitive to the starting frame. Bright or low-contrast images (especially white-dominant scenes) have a high chance of causing flickering, brightness fluctuations, or even a green tint/green screen effect, as shown below. This is a known behavior related to the generation process and not an issue with the checkpoint itself.

This LoRA was trained on a wide variety of data, which helps it capture diverse actions but can also introduce slight color instabilities in certain frames. Using a starting frame with more contrast or less white, or slightly lowering the LoRA weight, usually improves stability while preserving the intended actions.

This is my first Video LoRA trained with WAN2.2 I2V.
It still has a number of issues, but I may continue improving and updating it in the future (GPU permitting lol).