| license: openrail | |
| # miniSD-diff-loss-AD | |
| 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. |