Rui Zhao comments

Results 33 comments of


                                            Rui Zhao

Why are the inference results different from the results you posted?

There are two possible reasons. 1) Our torch version is different, which means that the same random seed may generate different noises. 2) The released model has been updated later....

Why are the inference results different from the results you posted?

Sorry for the late reply. I checked the results you provided and found that only the result of the "panda riding a bicycle" is abnormal while other results are the...

earlier checkpoints doing better than later checkpoints

I think it is abnormal to produce pure noise videos. Can you provide more details, such as the package versions and the configs you used?

About the training parameters of Spatial Lora: Recursive ones?

Sorry for the late reply, I was busy on another project. This is just a simple implementation for injecting one spatial LoRA for each video, and the multiple spatial LoRAs...

Does every new animation need to be retrained?

Yes, for new motions, you need to train the LoRAs again. This is a typical setting in the tuning-based customization task.

Errors in Lora : UnboundLocalError: local variable '_tmp' referenced before assignment`

Hi. What is your diffuser version?

Adding more data in the test folder .

This error seems like your video data is broken. Could you open the video file with a video browser?

Some confusion about the inconsistency between code and paper description

Hi. This trick is just for faster training. You can simply set the probability to 0%.

Some confusion about the inconsistency between code and paper description

In the forward process, setting lora_i = 0 makes sure the corresponding LoRA will not influence the outputs and the loss calculation. In the backward process, the corresponding optimizer will...

training time

This is abnormal. Can you post more information here?