[Windows] Installation Problems / Solutions

[MNeMoNiCuZ]

I mentioned elsewhere that the Readme is lacking some information about initialization and setup. So I figured I'd document my progress here so you know where I, as a sample user, got stuck.

1. Installation states: We rely on torch with CUDA installed correctly., but it doesn't mention either version. Nor order of installation.
2. Assuming order is important > Python > Cuda > Torch > Requirements. But please state the exact order for beginners.
3. Assuming python version. Installing with 3.10
4. Assuming Cuda 12.4. It's common and stable.
5. Assuming Pytorch 2.51 and torchvision 0.20.1 due to the help received by user Biaru mentioning that it's documented in the Huggingface. I cannot find it there, nor should I have to search for the information there. The code documentation should be here, where the code is :)
6. uv pip install torch==2.5.1 torchvision==0.20.1 torchaudio --index-url https://download.pytorch.org/whl/cu124
7. When installing requirements, I'm immediately met with this:

Starting installation from requirements.txt...
error: Expected `--hash`, found `"--no-build-isolation"` at requirements.txt:16:12

Which is of course, flash-attn --no-build-isolation
This is notoriously not installing on Windows. It's not clear if you are supporting Windows installation on the project page.

If I comment out the flash-attn, and retry installation, here's my next area of getting stuck:

Using Python 3.10.9 environment at: venv
Resolved 96 packages in 463ms
  x Failed to build `meshiki==0.0.5`
  |-> The build backend returned an error
  `-> Call to `setuptools.build_meta.build_wheel` failed (exit code: 1)

      [stdout]
      running bdist_wheel
      running build
      running build_py
      copying meshiki\__init__.py -> build\lib.win-amd64-cpython-310\meshiki
      running egg_info
      writing meshiki.egg-info\PKG-INFO
      writing dependency_links to meshiki.egg-info\dependency_links.txt
      writing requirements to meshiki.egg-info\requires.txt
      writing top-level names to meshiki.egg-info\top_level.txt
      reading manifest file 'meshiki.egg-info\SOURCES.txt'
      reading manifest template 'MANIFEST.in'
      adding license file 'LICENSE'
      writing manifest file 'meshiki.egg-info\SOURCES.txt'
      running build_ext
      building '_meshiki' extension
      "C:\Program Files (x86)\Microsoft Visual
      Studio\2022\BuildTools\VC\Tools\MSVC\14.40.33807\bin\HostX86\x64\cl.exe" /c /nologo /O2 /W3 /GL /DNDEBUG
      /MD -IC:\Users\user\AppData\Local\uv\cache\sdists-v9\pypi\meshiki\0.0.5\kgmA_ip9Y4tQwu7kEqWfb\src\include
      -IC:\Users\user\AppData\Local\uv\cache\builds-v0\.tmpFmInEX\lib\site-packages\pybind11\include
      -IC:\Users\user\AppData\Local\uv\cache\builds-v0\.tmpFmInEX\include
      -IC:\Python310\include -IC:\Python310\Include "-IC:\Program Files (x86)\Microsoft Visual
      Studio\2022\BuildTools\VC\Tools\MSVC\14.40.33807\include" "-IC:\Program Files (x86)\Microsoft
      Visual Studio\2022\BuildTools\VC\Auxiliary\VS\include" "-IC:\Program Files (x86)\Windows
      Kits\10\include\10.0.22621.0\ucrt" "-IC:\Program Files (x86)\Windows Kits\10\\include\10.0.22621.0\\um"
      "-IC:\Program Files (x86)\Windows Kits\10\\include\10.0.22621.0\\shared" "-IC:\Program
      Files (x86)\Windows Kits\10\\include\10.0.22621.0\\winrt" "-IC:\Program Files (x86)\Windows
      Kits\10\\include\10.0.22621.0\\cppwinrt" "-IC:\Program Files (x86)\Windows Kits\NETFXSDK\4.8\include\um"
      /EHsc /Tpsrc/bindings.cpp /Fobuild\temp.win-amd64-cpython-310\Release\src\bindings.obj /std:c++latest /EHsc
      /bigobj -std=c++17 -O3
      bindings.cpp
      C:\Users\user\AppData\Local\uv\cache\sdists-v9\pypi\meshiki\0.0.5\kgmA_ip9Y4tQwu7kEqWfb\src\include\meshiki/utils.h(105):
      warning C4267: 'argument': conversion from 'size_t' to 'const int', possible loss of data
      C:\Users\user\AppData\Local\uv\cache\sdists-v9\pypi\meshiki\0.0.5\kgmA_ip9Y4tQwu7kEqWfb\src\include\meshiki/utils.h(108):
      warning C4267: 'initializing': conversion from 'size_t' to 'int', possible loss of data
      C:\Users\user\AppData\Local\uv\cache\sdists-v9\pypi\meshiki\0.0.5\kgmA_ip9Y4tQwu7kEqWfb\src\include\meshiki/utils.h(109):
      ...
      ...
      ...
      warning C4244: 'initializing': conversion from 'float' to 'int', possible loss of data
      C:\Users\user\AppData\Local\uv\cache\sdists-v9\pypi\meshiki\0.0.5\kgmA_ip9Y4tQwu7kEqWfb\src\include\meshiki/mesh.h(960):
      warning C4267: 'argument': conversion from 'size_t' to 'int', possible loss of data
      [stderr]
      cl : Command line warning D9002 : ignoring unknown option '-std=c++17'
      cl : Command line warning D9002 : ignoring unknown option '-O3'
      error: command 'C:\\Program Files (x86)\\Microsoft Visual
      Studio\\2022\\BuildTools\\VC\\Tools\\MSVC\\14.40.33807\\bin\\HostX86\\x64\\cl.exe' failed with exit code 2

      hint: This usually indicates a problem with the package or the build environment.

So, meshiki is failing on Windows at least.
https://pypi.org/project/meshiki/ provided no further guide for Windows or anything.

If I comment out meshiki and install requirements, the rest installs fine.

Trying to launch py app.py:

(venv) C:\AI\3D\PartPacker>py app.py
Traceback (most recent call last):
  File "C:\AI\3D\PartPacker\app.py", line 25, in <module>
    from flow.model import Model
  File "C:\AI\3D\PartPacker\flow\model.py", line 21, in <module>
    from transformers import Dinov2Model
ModuleNotFoundError: No module named 'transformers'

Okay, not sure why transformers weren't part of requirements?
pip install transformers

Launching again:

Now it's downloading the models automatically to my hubcache directory.
So the pretrained models instruction wasn't necessary. I did this step the last time, but it failed with wget and others. You could also just have linked directly to the file, but since it auto-downloads, you could mention that it auto-downloads.
In fact, the gradio GUI auto-downloads even if you have the models in the pretrained directory. So this wget step does not actually help, not with the GUI at least.

After this download, the gradio GUI give me the URL and I can enter the server and it appears to work.

Feedback

The option of saving as OBJ/STL would be appreciated.

Fortunately these websites help converting from the GLB-format into something useful:
https://glb.ee/
https://convert3d.org/glb-to-obj/app

An option to auto-prune very tiny details would be nice.

Single triangle miniscule meshes inside other meshes are usually safe to delete. Maybe some algorithm that compares the extents of each mesh compared to the others, and ones that are much smaller than some other groups to be pruned?

Overall, great work with the model, it appears to work quite alright!

Commercial use when? It's a shame to not be able to use the model for other workflows or work. The licence kind of breaks the chain of potential of what could be done here.

[realstevewarner]

Thanks for this breakdown. I can confirm it works on Windows 10 using a 4090 and CUDA 12.4.

In short:

Install Python 3.10.
Install torch cuda 12.4 using:
pip install torch==2.5.1 torchvision==0.20.1 torchaudio --index-url https://download.pytorch.org/whl/cu124

Edit the requirements.txt file and eliminate meshiki and flash-attn. Neither are apparently necessary.
Install transformers (if it's not already installed) using pip install transformers

Run the app with py app.py

It will auto-download the needed models and then give you a URL for the gradio app.

On my 4090 it takes about 90 seconds to generate a model.

[ashawkey]

@MNeMoNiCuZ @realstevewarner Wow thanks for the testing!
I have removed flash-attn and meshiki in the default requirement.txt, and added transformers.
Those strange single-triangle parts seems to be a problem from trimesh, I will add a simple filtering for now.
Sorry that this model's license maybe hard to change, since its performance is not robust and maybe unsatisfactory. We'll try to improve it in future works, which may have a more open license!

[MNeMoNiCuZ]

Thanks for listening to the input.

Understandable about the licence of course, just figured I'd ask and express interest :)

[MNeMoNiCuZ]

I would still recommend updating the readme with the python/torch versions.

The instructions from @realstevewarner above is pretty good for a Windows installation guide.

Closing this issue will hide the information from people ^^

[ashawkey]

I see, will reopen and keep this issue!

[MNeMoNiCuZ]

Thanks.
If you'll consider accepting it I can make a pull request by adding it to the Readme instructions.

[ashawkey]

@MNeMoNiCuZ Sure that's better! Thanks!

[MNeMoNiCuZ]

Perfect! I added the instructions and made a pull request @ashawkey :
#11

@realstevewarner let me know if you have any suggested edits.

[ashawkey]

Merged! Thanks!

[joojie]

### Great job keep it up!

hey how's it going? k so I got the gradio to build on Win 11 + laptop 5090.

I had a lot if issues with the requirements not installing within the venv so I was hanging on simple things like:

"import opencv as cv2 module not found" and "import gradio as gr module not found"

Then it became rust and cargo hanging with the same issue. As a perpetual newbie I failed to prep with obviuos things like turning of virus protection and windows firewall. this seemed to cure the rustup package install which was timing out. the puzzling part was there was no notification of the timeout everything seemed to install. it wasn't until I tried the install using google gemini CLI that I was informed of the timeout issue ( CLI also suggested the firewall disable fix).

then after at CLI's suggestion "set Path=..." was used and then the app.py completed. it was a three day process.

When I started I thought there was an option to use multiple photos_ or maybe I was just imagining things in an overworked. state? Thats how long the install took - I actually think that in between beginning and completion the instructions and code were updated.
There was an error being thrown for "orch not found" which magically stopped appearing (maybe an error from the nightly version of torch +128 I was installing).

Thanks NVLabs/PartPacker team for a really nice package that works reliably with fairly high fidelity. I've only pushed the faces to 500k with all options maxed so far but so far so excellent.

### Great job keep it up!