Experimental Wan 2.2 i2v motion lora. Trained using 20 3-second long D-ART animation clips, all ass. Purpose is to see if an artist's animation style (not just the visual style) can be replicated via stable diffusion lora training.
Works best at 12 fps setting (generate ~37 frames with the export settings set to 12 fps for a ~3s video). All videos in the data set were three to four seconds long so it performs best when the exported video is 2 to 3 seconds long. If you notice that motions start looking stiff of off, try shaving a split second or so off of the video time (slightly less frames). It seems to perform more consistently and closer to the source material when the number of frames is in the 17 to 25 range.
Description
V2 Hig noise. Trained on 5 seconds of video instead of 3 seconds
FAQ
Comments (2)
The config.yaml and optimizer, where do they go?
Thanks
the config file is just a backup of the settings I nused to train these loras and the optimizer file is used by the trainer to enable continued training after a pause or sample generation. Neither are necisarry for inference. They're uploaded here mostly for my own convenience.