search

cure-lab / MagicDrive

[ICLR24] Official implementation of the paper “MagicDrive: Street View Generation with Diverse 3D Geometry Control”
https://gaoruiyuan.com/magicdrive/
GNU Affero General Public License v3.0
668 stars 40 forks source link

Bad controllability for box condition. #102

Open swc-17 opened 2 weeks ago

swc-17 commented 2 weeks ago

Hi, thanks for your great work. I build a model sharing the same camera/box encoder with Magicdrive, without map condition and BEVControlNet, and set all params in UNet trainable. After trained for 100 epochs, I found the generated images do not match with box condition well, for some boxes, the images do not have foreground objects at the corresponding location. Any suggestions on this? And I wonder how many epochs is the released model trained for? 350e as in the config? Thanks!

swc-17 commented 2 weeks ago

clipboard-image-1730862149

flymin commented 1 week ago

without map condition and BEVControlNet, and set all params in UNet trainable.

We never tried to train like this. You may adjust the parameters, like learning rate or batch size, to optimize the training process.

Typically, 100 epochs should be fine in most cases. The release model is trained with 350 epochs to optimize the results for quantitative evaluation.

github-actions[bot] commented 8 hours ago

This issue is stale because it has been open for 7 days with no activity. If you do not have any follow-ups, the issue will be closed soon.