Open Junoh-Kang opened 3 months ago
Answer For Q1:
We have tested controlling motions with box for non-fine-tuned model. In some cases, it is actually able to control the movement. However, you should set the controlling strength lower and test more samples, or the quality will be poor
Answer for I1:
In our implementation, the camera does influence the box sequence. You can look through pipeline_motionbooth
, the box motion and camera motion all are based on absolute coordinate system. So for example, if the box are set stable and camera moves to the left, the box in the video will move to the right, accordingly. The reason we do so is that latent shift is performed during the middle steps, where the approximate layout of objects are already fixed. If we do not move the object according to the shifted latent, the result quality might be lower. You can try it by remove the specific code in def latent_shift()
from pipeline_motionbooth
I think there is a misleading part in I1. The point is that when I use num_samples >1
, the bbox
values for first and second videos are different. I think it may not be an intended result. You can check this by drawing ani
every iteration. As a quick fix, I deepcopied initial bbox
.
You mean the bbox passed into the pipeline forward function to bbox.deepcopy()?
No, just at the inference.py, just before pipeline. ` bbox_ref = bbox.deepcopy() for i in num_samples: bbox = bbox_ref.deepcopy() pipe(bbox,...)
`
I think the problem might be the bbox value being changed in the pipeline. So passing bbox.deepcopy() into the pipeline might be ok. Your method is also ok.
Dear authors,
I have a question and an issue.
Q1. Is it possible to control motions with box for non-fine-tuned model? Or is only camera motion controllable? I experience strange objects when I use bbox option for pre-trained zeroscope.
I1. When I use
num_samples > 1
, it seems bbox value changes. It seemspipeline_motionbooth
changes bbox values with camera motion.