My first experiment with LoRa for CogVideoX1.5-5B-I2V
The quality of LoRa could suffer due to the following things:
1. Training on 45 videos of 13 frames with a fixed resolution of 720x480
2. A prompt of one word was used during training
3. Due to the peculiarity of the animation, very jerky movements are obtained.
I will try to retrain LoRa when Framer or some other good method of interpolating 2D video comes out.
Try experimenting with the speed of the video.
Here are the parameters that I used in the examples:
prompt: Slow motion, slow smooth dance, Live2D animation.
0.8 LoRa strenght
13 frames (Better use 53 since half of the video is static on 13)
25 step