Thank you for your great work!
I am trying to change camera poses at inference stage by inputting different camera_embedding to the pipeline - this gives me some unexpected results, so I tried to test a bit more about the camera_embedding parameter, and found some more unexpected results:
I firstly tried to input the default camera_embedding I found in mvdiffusion/pipelines/pipeline_mvdiffusion_image.py , and it works fine:
I wonder if this behaviour is expected? Does this mean that we are not supposed to change camera_embedding, and the output of inference step MUST be the 6 poses normal + the 6 poses rgb?
Thank you for your great work! I am trying to change camera poses at inference stage by inputting different
camera_embedding
to the pipeline - this gives me some unexpected results, so I tried to test a bit more about thecamera_embedding
parameter, and found some more unexpected results:I firstly tried to input the default
camera_embedding
I found in mvdiffusion/pipelines/pipeline_mvdiffusion_image.py , and it works fine:Code:
Result:![result_cameraemb_orig](https://github.com/xxlong0/Wonder3D/assets/45315408/f09ff7b0-81e3-4b44-a322-7582259516ba)
Then I tried to change the camera_embedding to the following, expecting to see the identical first image 12 times - but saw this different image:
Code:
Result:![result_cameraemb_0](https://github.com/xxlong0/Wonder3D/assets/45315408/e6c591bd-c487-4486-8f6f-71f289670cb2)
And I tried some other combinations and they are worse:
Code:
Result:![result_cameraemb_01](https://github.com/xxlong0/Wonder3D/assets/45315408/a9cc4fa8-16ca-49d7-b000-ab462c9c875d)
Code:
Result:![result_cameraemb_1](https://github.com/xxlong0/Wonder3D/assets/45315408/b4b9a562-cabf-451e-8631-302283a3254d)
I wonder if this behaviour is expected? Does this mean that we are not supposed to change
camera_embedding
, and the output of inference step MUST be the 6 poses normal + the 6 poses rgb?p.s., my full inference script is as follows:
Thank you so much! I look forward to your kind reply.