Releases · YanWenKun/Comfy3D-WinPortable

25 Jan 03:19

github-actions

r9-pt25

e90a23f

2025/01/25 Pre-release

Pre-release

This is an unstable pre-release, it is updated with ComfyUI-3D-Pack.

The instruction for installation is same as current README.

Added support for Hunyuan3D 2.0.
Dropped support for CharacterGen.
- CharacterGen depends on earlier versions of diffusers which is incompatible with Hunyuan3D-2.
- Maybe it's possible to eventually find a solution to make them compatible. But for now I don't get any clue.
Dropped support for CRM.
- CRM depends on earlier versions of open-clip-torch.
- I may add it back in stable releases, but for now let's focus on new models and new workflows.
Enabled Flash Attention by default.
- GPUs earlier than Ampere (30 series) may throw error when running some workflows.
- If you are using an old GPU and want to check out Hunyuan3D 2.0, try Hunyuan3D-2-WinPortable.

Versions Info

ComfyUI-3D-Pack: Jan 25, 2025 (2025/01/25)
ComfyUI 0.3.12 | PyTorch 2.5.1 | xFormers 0.0.29.post1 | CUDA 12.4 | Python 3.12.8

Assets 4

23 Jan 02:14

github-actions

r8-hunyuan3d2

30876e3

2025/01/22 Hunyuan3D-2 Special Pre-release

Pre-release

UPDATE

This sub-project has become a standalone project. Please visit the new repository and use the updated package:
本项目已移至独立仓库，请前往下载新版：
https://github.com/YanWenKun/Hunyuan3D-2-WinPortable

This is a special release for Hunyuan3D-2. It only runs the Hunyuan3D-2 Gradio app (official demo) and has nothing to do with ComfyUI. I put it here just for convenience.

Usage

Install MSVC and CUDA, see INSTRUCTIONS.
Download and extract Hunyuan3D2_WinPortable.7z.
Run the scripts 1-2-3.
Unlike Comfy3D-WinPortable, you don't need to edit these scripts, except for configuring internet proxies.
Open http://localhost:8080/
The output files are saved under Hunyuan3D-2\gradio_cache.

Notes

This bundle integrates the mmgp made by deepbeepmeep. Which means:
- 6GB of VRAM is sufficient for generating meshes only.
- 10GB of VRAM for generating meshes and textures.
- 24GB of system RAM is required.
The "Remove Background" feature uses rembg with default settings, which may output slightly white edges. Turn it off if your images already have clear alpha channels.
The "Text to 3D" feature is disabled by default. It consists of:
1. Text to Image
2. Image to 3D
  For experienced users, it might not be worth downloading another Text-to-Image model.
For GPUs earlier than Turing (RTX 20 series), a message "EP Error: Can't load TensorRT" will appear. And it will automatically fallback to CUDA, so you can safely ignore this message.

本次发布专门适配混元3D 2.0 的 Gradio App，与 ComfyUI 无关，放这里只是为了省事，毕竟用户群重合。

用法

安装 MSVC 和 CUDA，参照当前 Comfy3D-WinPortable 文档
下载并解压 Hunyuan3D2_WinPortable.7z ，把中文脚本复制到上级目录，无需修改脚本，按顺序 123 执行即可
启动完毕后，打开 http://localhost:8080/
生成的文件保存在 Hunyuan3D-2\gradio_cache 下

备注

本包整合了 mmgp，大幅减少显存需求，只需 6GB 显存就可生成网格，10GB 显存可生成网格与纹理。注意系统内存需要不低于 24GB。
移除背景（Remove Background）功能使用的是 rembg （默认配置），可能会生成轻微白边。如果你的图片本身已经抠干净背景了，就没必要勾选这个功能了。
文生 3D 的原理是先文生图，再图生 3D。对于有经验的用户而言，可能没有必要再下载一个文生图模型，因此默认不启用该功能。
对于 20 系之前的显卡，会提示 "EP Error: Can't load TensorRT"，忽略即可，程序找不到 Tensor 执行会自动改用 CUDA 执行。

THANKS TO:

https://github.com/Tencent/Hunyuan3D-2
deepbeepmeep 's VRAM optimization.
MrForExample 's work for Windows support.
meet-radek 's setup script for differentiable_renderer.
sdbds|青龙圣者 's setup script for differentiable_renderer.
RamiroG8k 's bugfix.

Thanks for all of you! It happens so quick and was truly amazing!

Versions Info
PyTorch 2.5.1 | CUDA 12.4 | Python 3.12.8

Modified repo used by this bundle: https://github.com/YanWenKun/Hunyuan3D-2
Build scripts for this bundle: https://github.com/YanWenKun/Comfy3D-WinPortable/tree/main/pt25-cu124-hunyuan3d2

Assets 3

21 Jan 19:09

github-actions

r7-pt25

bdcc3d9

2025/01/21 Latest

Latest

README | 使用说明

This is a maintenance release, with no major updates.

Updated Stable Fast 3D from Stability AI.
Migrated Python runtime from Python Embedded to Astral's python-build-standalone. Making it easier for compiling with Python header and lib files.

Versions Info

ComfyUI-3D-Pack: Dec 18, 2024 (2024/12/18)
ComfyUI 0.3.12 | PyTorch 2.5.1 | xFormers 0.0.29.post1 | CUDA 12.4 | Python 3.12.8

Assets 4

21 Jan 09:46

YanWenKun

r7-wheels

36f099f

Wheel Files for pt25-cu124 Pre-release

Pre-release

This is a functional release for uploading compiled wheel files.
You may safely ignore this release.

Instructions for compiling

Assets 4

30 Dec 22:25

github-actions

r6-pt25

41e0ca4

2024/12/31

README | 使用说明

Changes

Added support for TRELLIS and Hunyuan3D-1.
Integrated Stable Fast 3D from Stability AI.
Newly compatible with 5 workflows.
Pre-installed ComfyUI-Manager. Please don't use it to update 3D-Pack!

Versions

ComfyUI-3D-Pack: Dec 18, 2024 (2024/12/18)
PyTorch 2.5.1
CUDA 12.4
Python 3.12

Assets 4

30 Dec 11:17

github-actions

r5-pt25

ec8cfbb

2024/12/30 PT25 Preview Pre-release

Pre-release

Dropped support for TRELLIS gradio app (due to conflict of diff-gaussian-rasterization).
Improved compatibility. Most workflows should work now (except CRM_InstantMesh).
TRELLIS workflow is still testing. Please let me know if it works for you or not.
Pre-installed ComfyUI-Manager (please do not use it to update 3D-Pack).

I'm still testing, debugging and writing docs. But this build is almost ready so I released it as a preview.

Update:

The reason why TRELLIS workflow isn't working on some GPU is because SpConv will compile CUDA kernel in runtime for GPU archs not in pre-built list, and the compiler can't find <.h> headers, which is an old issue for Python Embedded. This is fixed in next stable release so you can just ignore this one.

Assets 4

20 Dec 05:05

github-actions

r4-pt25

9825b4e

2024/12/20 TRELLIS Special Pre-release

Pre-release

This is a special release for TRELLIS. I only tested the TRELLIS gradio app (official demo) ~~and the TRELLIS workflow of 3D-Pack~~.

The usage is generally the same as in the current README, with some key points I need to mention:

Make sure to install CUDA Toolkit 12.4.1. Other versions may not work.
The new script run_trellis_demo.bat will start the TRELLIS official gradio demo, similar to the one available on HuggingFace Space.
Don't forget to edit the scripts before starting them.
ComfyUI-Manager is pre-installed in this release. DO NOT use it to update 3D-Pack!

Note: TRELLIS needs at least 16GB of VRAM. On Windows, it can use CPU RAM when GPU VRAM is not enough, but it will be very slow.

Update: You can safely update the TRELLIS gradio app by running git pull. No further steps are needed.

本次发布只测试了官方 demo ~~与 TRELLIS 工作流~~ 。使用方法参照当前的说明文档即可，额外注意：

一定要安装 CUDA Toolkit 12.4.1，其他版本很可能不兼容。
新加了一个启动脚本 run_trellis_demo_cn.bat，可以启动 TRELLIS 官方 demo，即 HuggingFace Space 同款。
别忘了先运行编译安装脚本，也别忘了先编辑脚本再启动。
本次预装了 ComfyUI-Manager，但千万别用它更新 3D-Pack，免得依赖项炸锅。

额外：TRELLIS 需要至少 16GB 显存。Windows 驱动会在显存不足时调用内存，让 12G 显存的显卡也能跑，但速度极慢。

更新：你可以通过 git pull 来更新 TRELLIS gradio app 到最新版本。新版的依赖项无变动，因此无需更多步骤。

Versions info
Python 3.12.8
CUDA 12.4
PyTorch 2.5.1
xFormers 0.0.28.post3

Assets 4