fudan-generative-vision
Finetune on costom dataset
Thanks for opensource such great work. But I found the result on some full body dataset is not good enough, relatively speaking. I want to try fine-tune the stage 2 using some full body pose dataset like zjumocap, snapshot and etc. But I found when I loaded your weight and trained I found the validation result is like noise
I am pretty sure I load all your weight first. And it is weird when I didn't load motion_module, the result seems ok.
This is indeed strange. If you haven't changed the network structure, I recommend checking if there is an issue with the motion module weights being corrupted. It might be a good idea to re-download the motion module.
This is indeed strange. If you haven't changed the network structure, I recommend checking if there is an issue with the motion module weights being corrupted. It might be a good idea to re-download the motion module.
It seems I changed the frames number. If I keep the original frame number, it becomes normal