A semi-portable package dedicated for running ComfyUI-3D-Pack on Windows.
This build, pt24-cu121, is based on PyTorch 2.4.0 with CUDA 12.1,
targeting the
Oct 18, 2024
version of ComfyUI-3D-Pack. Workflows that are "too new" or "too old" are not supported in this build.
I have tested all the example workflows on a fairly-old TITAN Xp GPU (12GB VRAM). The results are as follows:
| Workflow | Status | Screenshot | Memo |
|---|---|---|---|
CharacterGen_Vanilla |
🆗 |
||
CharacterGen_to_Unique3D |
🆗 |
||
CharacterGen_InstantMesh |
❓ |
Out of memory at "InstantMesh Reconstruction Model" node. |
|
CRM MVDiffusion_to_Craftsman |
🆗 |
||
CRM MVDiffusion_to_Unique3D |
🆗 |
||
CRM_Img_to_Multi-views_to_CCMs_to_Mesh |
🆗 |
||
CRM MVDiffusion_to_InstantMesh |
❌ |
PyTorch version incompatible. |
|
CRM_T2I_V2 |
❓ |
||
Unique3D_All_Stages |
🆗 |
As well as Unique3D stage 0 to 3. |
|
Unique3D_with_Era3D_All |
❌ |
PyTorch version incompatible. |
|
[FlexiCubes] Multi-View Depth plus Optional Normal maps to 3D Mesh |
🆗 |
||
Era3D_to_InstantMesh |
❌ |
PyTorch version incompatible. |
|
Multi-View-Images_to_Instant-NGP_to_ DiffRast_to_3DMesh |
❌ |
Throwing ('NoneType' object has no attribute 'ray_aabb_intersect') at "Instant NGP" node. |
|
MVDream_to_Large_Multiview_Gaussian_ Model_to_3DGS and_3DMesh |
❌ |
Throwing ('NoneType' object has no attribute 'ray_aabb_intersect') at "Convert 3DGS to Mesh with NeRF and Marching Cubes" node. |
|
Render_Mesh_and_3DGS_Example |
🆗 |
||
StableFast3D_to_Mesh |
🆗 |
||
Triplane_Gaussian_Transformers_to_ 3DGS_and_3DMesh |
❌ |
Throwing ('NoneType' object has no attribute 'ray_aabb_intersect') at "Convert 3DGS to Mesh with NeRF and Marching Cubes" node. |
|
TripoSR_to_Mesh |
🆗 |
The "easy" model that always works. |
|
Using_Stack_Orbit_Camera_Poses_ to_conditioning_StableZero123 |
🆗 |
||
Wonder3D_to_Large_Multiview_ Gaussian_Model_to_3DGS and_3DMesh |
❌ |
Throwing ('NoneType' object has no attribute 'ray_aabb_intersect') at "Convert 3DGS to Mesh with NeRF and Marching Cubes" node. |
|
Zero123Plus_to_InstantMesh |
🆗 |
Few things I want to mention first, that may save you A LOT of time…
-
If you’re looking for Stable Fast 3D, which generate 3D-mesh from single image, just use StableFast3D-WinPortable. It does only one thing and does it well.
-
Everything in Comfy3D works more smoothly on Linux, as well as WSL2. Most 3D workflows are research-oriented and far from production-ready. By installing & running them you’re essentially replicating the development environment used by researchers. Many components need to be compiled before running, and this is exactly where Linux excels. Check out my Docker image if you’re familiar with containers.
-
Know your GPU’s CUDA arch number. For example, RTX 2080 is
7.5, RTX 3090 is8.6and RTX 4090 is8.9. You can find the number in this article.
1.1. Install CUDA Toolkit 12.1.1 or CUDA Toolkit 12.4.1. Both are fine, choose 12.4.1 if you have other newer ComfyUI instances.
-
Only Libraries and Compilers are needed to install.
-
If your GPU driver is outdated (<530 for CUDA 12.1, <550 for CUDA 12.4), you can install the driver provided by the CUDA Toolkit installer. Otherwise, you don’t need to downgrade the driver version.
-
During the setup, select "Desktop development with C++" workload.
2.1. Download the package files from the release page:
-
Note that there are multiple files (compressed in volumes) that need to be downloaded. Extract it to a convenient place.
2.2. Edit !FIRST_RUN-compile-install.bat
You can see the line:
set TORCH_CUDA_ARCH_LIST=6.1+PTX
Edit it according to your GPU arch. For example, for RTX 4060 Ti, it should be:
set TORCH_CUDA_ARCH_LIST=8.9
2.3. Edit run.bat
You can see the line:
set TORCH_CUDA_ARCH_LIST=5.2+PTX;6.0;6.1+PTX;7.5;8.0;8.6;8.9+PTX
Edit it like you just did in the last step.
There are few more settings available in this start script, such as:
-
Set proxy
-
Do not open browser after starting
Each line of configuration has a description above it. If needed, uncomment the according line (remove `rem ` at the beginning), edit it and save the file.
3.1. The ComfyUI-Manager and AIGODLIKE-ComfyUI-Translation are installed but disabled.
If you really want to use them, in ComfyUI\custom_nodes folder, rename them, remove the trailing .disabled so they will be loaded on start.
-
ComfyUI-Manager may damage the dependencies of Comfy3D, so it’s disabled by default.
-
AIGODLIKE-ComfyUI-Translation is for multilingual translation of ComfyUI.
3.2. If you want to use Sandboxie to isolate this instance, the following steps should be executed in a sandbox.
In addition, I recommend to set the program folder (Comfy3D_WinPortable) to "Full Access" under "Sandbox Settings - Resource Access", to improve disk I/O.
4.1. Run !FIRST_RUN-compile-install.bat. Wait until it finish installation. If it fails, simply try again.
4.2. If you are going to run Unique3D workflows, run download-models-for-Unique3D.bat.
4.3. Run run.bat.After starting, the program will automatically open a browser, or you can visit: http://localhost:8188/
4.4. In the left sidebar, you can find the "Workflows" button, the example workflows are listed there. And the example input files are already in the input folder. Open a workflow and click "Queue" at bottom of the page to run the workflow.
It is not recommended to update any component of this package, whether it’s Python, ComfyUI or Custom Nodes ("If it works, don’t fix it").
But if you really need to update ComfyUI and Custom Nodes (except 3D-Pack), there’s a script to do that. Run this in your Git Bash:
bash force-update-all.sh
This script will force git-pull all the repositories.
Note that ComfyUI\custom_nodes\ComfyUI-3D-Pack folder is not a git repository, so it won’t be affected by this script.
Just like the FIRST_RUN script did,
the reinstall-deps-for-3d.bat script will attempt to compile and install dependencies of 3D-Pack.
You need to edit this file and set TORCH_CUDA_ARCH_LIST as well.
There are few differences from the FIRST_RUN, this script will:
-
Download the latest source repos, not the targeted versions. This may help with compatibility issues when the FIRST_RUN doesn’t work.
-
Additionally compile-install
kiuikitandnvdiffrast. In my tests, it’s not necessary to do so in the FIRST_RUN. I put them here for fail-safe. -
Keep temporary files and save artifacts (.whl files) in the working folder
tmp_build, allowing you to reuse them. Keep in mind they are not essentially portable.