Releases: YanWenKun/ComfyUI-Windows-Portable
Releases · YanWenKun/ComfyUI-Windows-Portable
2026/01/26
Changes | 改动
- Upgrade to PyTorch 2.10.0.
- PyTorch 升级到 2.10.0。
Note | 提示
- In order to utilize SageAttention 2++, you need to install:
- CUDA Toolkit with Libraries and Compilers.
- MSVC with C++ workload.
- 如要利用 SageAttention 2++，需安装：
- CUDA Toolkit，安装时选择 Libraries 与 Compilers。
- MSVC，安装时选择“桌面C++开发”。
Versions Info | 版本信息
What's Next | 新版展望
- This is the last CUDA 12.8-based build. Future releases will transition to CUDA 13.0.
- ComfyUI is moving fast. New comfy-kitchen defaults to CUDA 13.0.
- 这是最后一版基于 CUDA 12.8 的发布，下次发布将基于 CUDA 13.0。
- ComfyUI 一直很追新，新出的 comfy-kitchen 默认 CUDA 13.0。
2026/01/22 CUDA 13.0
2025/12/23 CUDA 13.0
Changes since the last experimental release:
相对 最近一版试验包 的变动：
- Add Custom Node:
- 添加自定义节点：
Versions Info | 版本信息
2025/12/08 CUDA 13.0
In English:
- This is an experimental release. There are no significant changes since the last experimental release — only routine updates.
Difference
List of accumulated changes, compare to CUDA 12.8 stable release:
- CUDA 12.8 ➔ 13.0
- Python 3.12 ➔ 3.13 (with GIL)
- NumPy 1 ➔ 2
- Add FlashAttention
- Drop xFormers
- Drop node: ReActor
- Drop node: Jovimetrix
- Drop node: AKatz Nodes
- Drop node: InstantID
- Drop node: PuLID
- Drop node: FaceAnalysis
- Drop node: ComfyUI-SAM2
Note
- In order to utilize SageAttention 2++, you need to install:
- CUDA Toolkit with Libraries and Compilers.
- MSVC with C++ workload.
Versions Info
中文说明：
- 这是一版实验包，核心组件较新。与 上一版实验包 相比，无变更，仅日常维护。
与稳定版的差异
相比于 CUDA 12.8 分支的稳定版 ，积累的改动有：
- CUDA 12.8 ➔ 13.0
- Python 3.12 ➔ 3.13 （非自由线程，带 GIL）
- NumPy 1 ➔ 2
- 添加 FlashAttention
- 移除 xFormers
- 移除节点 ReActor
- 移除节点 Jovimetrix
- 移除节点 AKatz Nodes
- 移除节点 InstantID
- 移除节点 PuLID
- 移除节点 FaceAnalysis
- 移除节点 ComfyUI-SAM2
提示
- 如果你要用到 SageAttention 2++，需要安装：
- CUDA Toolkit，安装时选择 Libraries 与 Compilers。
- MSVC，安装时选择“桌面C++开发”。
版本信息
2025/12/07
Changes | 改动
- Update to PyTorch 2.9.1.
- PyTorch 更新到 2.9.1。
Note | 提示
- In order to utilize SageAttention 2++, you need to install:
- CUDA Toolkit with Libraries and Compilers.
- MSVC with C++ workload.
- 如要利用 SageAttention 2++，需安装：
- CUDA Toolkit，安装时选择 Libraries 与 Compilers。
- MSVC，安装时选择“桌面C++开发”。
Versions Info | 版本信息
- ComfyUI 0.3.76
- Python 3.12.12
- CUDA 12.8
- PyTorch 2.9.1
- xFormers 0.0.33.post2
2025/11/27
Changes | 改动
- No changes, just updates, to have out-of-box support for FLUX.2 and Z Image.
- 无改动，仅更新，以开箱支持 FLUX.2 和 造相 。
Hint | 提示
- In order to utilize SageAttention 2++, you need to install:
- CUDA Toolkit with Libraries and Compilers.
- MSVC with C++ workload.
- 如要利用 SageAttention 2++，需安装：
- CUDA Toolkit，安装时选择 Libraries 与 Compilers。
- MSVC，安装时选择“桌面C++开发”。
Versions Info | 版本信息
- ComfyUI 0.3.75
- Python 3.12.12
- CUDA 12.8
- PyTorch 2.9.0
- xFormers 0.0.33.post1
2025/11/14 PyTorch 2.9.1
- This is a sibling release of the recent experimental release.
- This Package does not come with xFormers. Instead, it come with the latest PyTorch 2.9.1, which fixed several memory/performance issues.
- If you only use ComfyUI to generate images, or if you are familiar with SpargeAttention/SageAttention/FlashAttention for video generation, this package is suitable for the scenario.
- 本包与 最近一版试验包 基本相同。
- 区别只是本包没有预装 xFormers，并得以预装最新的 PyTorch 2.9.1，该版修复了 2.9.0 的若干显存与性能问题。这些问题并未造成普遍影响，但如果你碰巧遇到，则本包便有了意义。
- 此外，如果你只用 ComfyUI 生成图片，或者你熟悉运用 SpargeAttention/SageAttention/FlashAttention 生成视频，也可考虑本包。
Versions Info | 版本信息
2025/11/14 New Attentions
-
This is a follow-up of the last experimental release. This package choose latest components over compatibility for Custom Nodes.
-
作为上一版激进升级的后续，轻装上阵的新版预装了新的性能组件。
Changes | 改动
-
Add performance packages | 添加性能库
-
Add xFormers back | xFormers 回归
Add Custom Nodes | 新增预装节点
Versions Info | 版本信息
- ComfyUI 0.3.68
- Python 3.13.9
- CUDA 13.0
- PyTorch 2.9.0
- xFormers 0.0.33.post1
2025/11/14
- This is a compatibility-focused release, preserving all Custom Nodes from the last stable release onward.
- 这是一版旨在兼容性的发布，与 上一稳定版 相比无任何组件删减。
Changes | 改动
- Upgrade to PyTorch 2.9.0. | 更新 PyTorch
- Add performance packages | 添加性能库
- SpargeAttention
- SageAttention
- Add triton-windows as dependency | 添加依赖项
Add Custom Nodes | 新增预装节点
Hint | 提示
- In order to utilize SageAttention 2++, you need to install:
- CUDA Toolkit with Libraries and Compilers.
- MSVC with C++ workload.
- 如要利用 SageAttention 2++，需安装：
- CUDA Toolkit，安装时选择 Libraries 与 Compilers。
- MSVC，安装时选择“桌面C++开发”。
Versions Info | 版本信息
- ComfyUI 0.3.68
- Python 3.12.12
- CUDA 12.8
- PyTorch 2.9.0
- xFormers 0.0.33.post1
2025/11/01 Python3.13 NumPy2
- This is an experimental release. It includes breaking changes and deprecations.
- 这是一版激进升级的实验性版本，若干老组件失效。
Changes | 改动
- Upgraded to Python 3.13 (with GIL).
- Transitioning to free-threaded Python for this project is not feasible now.
- Upgraded to NumPy 2.
- Dropping NumPy 1 breaks nodes. It's painful to upgrade but it's more painful to not upgrade.
- Removed temporary added SageAttention.
- 升级到 Python 3.13（带 GIL）。
- 目前的生态完全不支持迁移到自由线程模式，除非只使用基本的 PyTorch。
- 升级到 NumPy 2。
- 一些节点还在死守 NumPy 1，但另外一些组件已经完全面向 NumPy 2 了，在此局面下只能取新舍旧。
- 移除临时添加的 SageAttention。
Removed Custom Node | 移除预装节点
- ReActor
- Jovimetrix
- AKatz Nodes
- InstantID
- PuLID
- FaceAnalysis
- ComfyUI-SAM2
- The dependency SAM2 from FAIR was kept.