Type | |
Stats | 545 0 158 |
Reviews | (77) |
Published | Apr 1, 2025 |
Base Model | |
Training | Steps: 130,000 Epochs: 8 |
Trigger Words | See workflow and description |
Training Images | Download |
Hash | AutoV2 1A074EC62A |
This is an experimental image to image LORA for Wan. I have also added a workflow because I probably spent 12h+ simply finding the best parameters to make it usable. The expression changer works well, and direction top, bottom, forward work fine, but unfortunately the left and right direction are not very consistent so might as well treat them as random. At least the generations are very fast because we are only generating 5 frames.
The workflow and prompt structure are present in the Zip file.
If you're feeling extra brave, you may try changing the parameters and see if you can make it work better.
These are the actual training prompts:
"The {man|woman} is standing in front of a green screen, {he|she} looks {forward|up|down|to the right_30|to the right_60|to the left_30|to the left_60} with a {barely|regular|very} {angry|contempt|disgusted|fear|happy|neutral|sad|surprised} emotion"
If you go this route, take into account left and right directions were inverted during training.