-
Notifications
You must be signed in to change notification settings - Fork 559
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Add fast high resolution marching cubes up to 1024^3. #68
base: main
Are you sure you want to change the base?
Conversation
Move density to CPU, further improve VRAM usage.
Revert to official marching cube algorithm.
2048^3 takes 15GB VRAM, 3.5 minutes. just change 1024 to 2048. But the mesh is to big for web GUI to display, you can still download the OBJ. |
只是Resolution数字更大了,还是mesh效果也明显更好了? |
大佬,可以放一下这样修改分辨率的结果嘛,修改之后有明显提升嘛mesh效果 |
分辨率大了肯定效果更好,面很锐。但是,这个本质还是插值,原本的 Triplanar分辨率较低的 @yuedajiong |
|
@thatname 谢谢大神 按照我有限理解,所以要提高质量:
|
哈哈,我不算大神啦。 速度 |
感觉我说的还是不对,不一定说从2D模型来生成3D就一定要逐个走优化,应该是能直接建立2D->3D映射的,这个过程是确定的,训练出来就行。 |
厉害的 谢谢大神 |
大神,意思是目前生成质量效果最好的还是SDS训练的方法嘛,SV3D这种其次,然后才是triposr这种直接出3D模型的方法嘛,可我测下来为啥觉得triposr的生成质量效果是最好呢,如果有比它好的,大神可以说一下对应的开源项目的名字嘛?万分感谢!!! |
大佬,这个生成的.obj的mesh确实分辨率可以很大,但是生成的mesh好像有的是non-manifold mesh,在用该mesh作别的任务时就会报该错误,大佬有啥解决方法嘛 |
我估计主要是因为这个分块生成,导致mesh是非连续的。直接合并坐标相同的顶点处理一下就好了。 |
可能TripoSR 训练还是不足,见过的物体还是太少,很多没见过的就脑补。不如图片生成 和视频生成模型见多识广。当然了,很多它见过的物体生成很不错。但这个问题是无解的,网上视频和图片是海量的,3D dataset 太有限 |
TypeError: extract_mesh() got multiple values for argument 'resolution' |
Is marching cubes now possible above 320? I just tried with TripoSR and got his error |
用不了啊?TypeError: extract_mesh() got multiple values for argument 'resolution' |
This is caused by a mismatch between function def and function calls Both these instances have scenecodes, some boolean variable, and then resolution. Fix is to remove the boolan in the calls. |
Changes:
This version can generate 1024^3 on 3090 in 33 seconds, use only 13GB VRAM.
Actually this version is block based, only a 128^3 block is processed at each step.
So 2048^3 is doable but unnecessary.
这个分支修改了算法,每次处理 128^3 的小块,还优化了显存占用。1024 的 mesh 在 3090 上半分钟跑完,256^3低分辩率也比之前快。