Output of i2V model currently not satisfying
The steps needs to be higher. I've found that the motion/quality doesn't get close to any closed-source ones until about 100 steps. It also helps marginally to not use any speedups like torch or SageAttention as they introduced very slight jolts in the motion.
Thx, but 100 steps will take ages on my 4070. I am not sure yet about the generation problem issue it seems the input picture is the key point which means the training/finetune of the model are the main problem. Same setting 1 turns out very good the other output is just garbage. 30 steps gentime around 2600 secs
Prompt: FPS-24, a woman with long, wavy blonde hair, wearing a brown blazer over a black collared shirt. She is standing indoors, possibly in a professional setting, as suggested by her attire and the blurred background.
Closing this because Kijais new workflow for native nodes works fine.