This is motion module V2 fine-tuning experiment that tries two novel things.

One it uses miniSD as a base model when training. This leads to increased detail and clarity when generating. Two, it uses a loss function that ensures the frame diffs are similar between the target and predicted. This reduces high frequency movement, e.g. flicker.

The model might have some issues. It doesn't always produce as much motion as V2 and can require more steps for context blending.