gsgen3d / gsgen

[CVPR 2024] Text-to-3D using Gaussian Splatting
https://arxiv.org/abs/2309.16585
MIT License
761 stars 42 forks source link

训练迭代14500次,为什么效果很差还存在多面问题? #34

Open jim-1ee opened 8 months ago

jim-1ee commented 8 months ago

WechatIMG127

throb081 commented 6 months ago

您好,我这边也想跑跑这个代码,但是我出了点问题,想咨询一下可以么,有偿。主要是这个BUG: DeprecationWarning: Type google.protobuf.pyext._message.ScalarMapContainer uses PyType_Spec with a metaclass that has custom tp_new. This is deprecated and will no longer be allowed in Python 3.14.

jim-1ee commented 5 months ago

你好,我没有遇到过,祝顺利~

------------------ 原始邮件 ------------------ 发件人: "gsgen3d/gsgen" @.>; 发送时间: 2024年3月10日(星期天) 上午9:40 @.>; @.**@.>; 主题: Re: [gsgen3d/gsgen] 训练迭代14500次,为什么效果很差还存在多面问题? (Issue #34)

您好,我这边也想跑跑这个代码,但是我出了点问题,想咨询一下可以么,有偿。主要是这个BUG: DeprecationWarning: Type google.protobuf.pyext._message.ScalarMapContainer uses PyType_Spec with a metaclass that has custom tp_new. This is deprecated and will no longer be allowed in Python 3.14.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

Deaddawn commented 4 months ago

WechatIMG127

same problem here

Deaddawn commented 4 months ago

您好,我这边也想跑跑这个代码,但是我出了点问题,想咨询一下可以么,有偿。主要是这个BUG: DeprecationWarning: Type google.protobuf.pyext._message.ScalarMapContainer uses PyType_Spec with a metaclass that has custom tp_new. This is deprecated and will no longer be allowed in Python 3.14.

Just a warning, should be ok for your training process

jimmychuang95 commented 2 months ago

Same here, 15000 iterations still occur severe Janus problem!

viridityzhu commented 1 month ago

Same here. tried several prompts, such as corgi, panda, and man in a suit. All occurred severe Janus problem

--- update --- the previous poor results were because when executing this command:

python main.py --config-name=base prompt.prompt="<prompt>"

the prompt for point-E was set to "a human face" as defined in base.yaml.

Instead, after specifying init.prompt="<point-e prompt>", I got good results without Janus problem.

BTW, it seems that the default configuration (conf/base.yaml) does not enable full functions, particularly the compactness-based densification and pruning, which are major contributions of the paper. Why are they not enabled? Wouldn't this lead to sub-optimal results?