error on generated normals after training

xxlong0 / Wonder3D

Single Image to 3D using Cross-Domain Diffusion for 3D Generation

https://www.xxlong.site/Wonder3D/

GNU Affero General Public License v3.0

4.49k stars 351 forks source link

error on generated normals after training #164

Open jingyu198 opened 2 months ago

jingyu198 commented 2 months ago

Very wonderful work and amazing demos!

I am trying to use my own data to train the model for stage 1 and use your weights as pretrained model. The only changes include pretrained model path and training data path.

However, when testing using test_mvdiffusion_seq.py, the normals seem almost the same as the RGB images. why does this happen? 1713190685452 res

jingyu198 commented 2 months ago

when training, there is a bug indicating that "can not find mask file", thus I set read_mask to False, does this lead to this unreasonable result?

xxlong0 commented 2 months ago

Fixed a severe training bug. The "zero_init_camera_projection" in 'configs/train/stage1-mix-6views-lvis.yaml' should be False. Otherwise, the domain control and pose control will be invalid in the training.