ComfyUI Spark-H3 v0.1.2

中文说明

本次发布新增经过验证的 Windows 安装包，并更新了内附工作流和安装文档。

支持环境

• ComfyUI 0.38.x 和 0.39.x
• Python 3.12+
• PyTorch CUDA 13.0
• Linux 或 Windows x86_64
• NVIDIA SM89：GeForce RTX 40 系
• NVIDIA SM120：GeForce RTX 50 系及 RTX PRO 5000/6000 Blackwell

安装器会根据当前 ComfyUI 环境自动选择对应后端：

• ComfyUI 0.38.x：comfy-kitchen 0.2.36
• ComfyUI 0.39.x：comfy-kitchen 0.2.37

本次发布暂不提供 CUDA 12.8 安装包。

安装包

请同时根据操作系统和 GPU 架构选择对应文件：

• ComfyUI-Spark-H3-0.1.2-linux-x86_64-sm89-cu130.zip
• ComfyUI-Spark-H3-0.1.2-linux-x86_64-sm120-cu130.zip
• ComfyUI-Spark-H3-0.1.2-windows-x86_64-sm89-cu130.zip
• ComfyUI-Spark-H3-0.1.2-windows-x86_64-sm120-cu130.zip

Windows SM89 和 SM120 安装包均已在对应目标平台完成验证。

主要变更

• 新增 Windows x86_64 CUDA 13.0 的 SM89 和 SM120 安装包。
• 示例工作流恢复使用 ComfyUI 默认节点显示名称。
• SM89 安装包中的工作流使用 SM89 节点，SM120 安装包中的工作流使用 SM120 节点。
• LBH 工作流仅在三步高分辨率阶段应用 Spark，并关闭 dense warmup。
• 更新中英文安装和使用文档。

安装方法

根据操作系统和 GPU 架构选择对应 ZIP，并将其解压到 ComfyUI/custom_nodes。

Linux

cd /path/to/ComfyUI/custom_nodes
unzip /path/to/ComfyUI-Spark-H3-0.1.2-linux-x86_64-<sm89-or-sm120>-cu130.zip
/path/to/ComfyUI/.venv/bin/python ComfyUI-Spark-H3/install.py

Windows

Expand-Archive `
  -Path C:\path\to\ComfyUI-Spark-H3-0.1.2-windows-x86_64-<sm89-or-sm120>-cu130.zip `
  -DestinationPath C:\path\to\ComfyUI\custom_nodes

& "C:\path\to\ComfyUI-Python\python.exe" `
  "C:\path\to\ComfyUI\custom_nodes\ComfyUI-Spark-H3\install.py"

请使用启动 ComfyUI 的同一个 Python。ComfyUI 便携版通常使用 python_embeded\python.exe，虚拟环境安装通常使用 .venv\Scripts\python.exe。安装完成后请重启 ComfyUI。

请勿混用不同操作系统、GPU 架构或 CUDA 版本的安装包和后端 wheel。

安装包不包含 ComfyUI 本体、模型权重、可选 LoRA 或第三方工作流节点。

SHA256

校验值见 Release 附件中的 SHA256SUMS.txt。

---

ComfyUI Spark-H3 v0.1.2

This release adds validated Windows packages and updates the bundled workflows and installation documentation.

Supported environments

• ComfyUI 0.38.x and 0.39.x
• Python 3.12+
• PyTorch CUDA 13.0
• Linux or Windows x86_64
• NVIDIA SM89: GeForce RTX 40 series
• NVIDIA SM120: GeForce RTX 50 series and RTX PRO 5000/6000 Blackwell

The installer automatically selects the matching backend:

• ComfyUI 0.38.x: comfy-kitchen 0.2.36
• ComfyUI 0.39.x: comfy-kitchen 0.2.37

CUDA 12.8 packages are not included in this release.

Release packages

Choose the archive matching both your operating system and GPU architecture:

• ComfyUI-Spark-H3-0.1.2-linux-x86_64-sm89-cu130.zip
• ComfyUI-Spark-H3-0.1.2-linux-x86_64-sm120-cu130.zip
• ComfyUI-Spark-H3-0.1.2-windows-x86_64-sm89-cu130.zip
• ComfyUI-Spark-H3-0.1.2-windows-x86_64-sm120-cu130.zip

The Windows SM89 and SM120 packages have been validated on their target platforms.

Changes

• Added Windows x86_64 CUDA 13.0 packages for SM89 and SM120.
• Bundled workflows now use the default ComfyUI node display names.
• SM89 packages contain SM89 workflow nodes; SM120 packages contain SM120 workflow nodes.
• The LBH workflow applies Spark only to the three-step high-resolution stage and uses zero dense warmup.
• Updated English and Chinese installation documentation.

Installation

Choose the ZIP matching your operating system and GPU architecture, then extract it under ComfyUI/custom_nodes.

Linux

cd /path/to/ComfyUI/custom_nodes
unzip /path/to/ComfyUI-Spark-H3-0.1.2-linux-x86_64-<sm89-or-sm120>-cu130.zip
/path/to/ComfyUI/.venv/bin/python ComfyUI-Spark-H3/install.py

Windows

Expand-Archive `
  -Path C:\path\to\ComfyUI-Spark-H3-0.1.2-windows-x86_64-<sm89-or-sm120>-cu130.zip `
  -DestinationPath C:\path\to\ComfyUI\custom_nodes

& "C:\path\to\ComfyUI-Python\python.exe" `
  "C:\path\to\ComfyUI\custom_nodes\ComfyUI-Spark-H3\install.py"

Use the same Python interpreter that starts ComfyUI. For ComfyUI Portable, this is normally python_embeded\python.exe; for a virtual environment, it is normally .venv\Scripts\python.exe. Restart ComfyUI after installation.

Do not mix packages or backend wheels from different operating systems, GPU architectures, or CUDA versions.

Model weights, optional LoRAs, and third-party workflow nodes are not included.

ComfyUI Spark-H3 v0.1.1：Linux CUDA 13.0

ComfyUI Spark-H3 v0.1.1：Linux CUDA 13.0

本版本提供面向 ComfyUI 原生 MiniMax-H3 模型的 Spark-H3 块稀疏注意力插件，并包含针对不同 NVIDIA GPU 架构预编译的 CUDA 13.0 后端。

主要功能

• 支持 NVIDIA SM89：GeForce RTX 4090。
• 支持 NVIDIA SM120：GeForce RTX 50 系列及 RTX PRO 6000 Blackwell。
• 支持 ComfyUI 0.38.x 和 0.39.x。
• 内置对应 GPU 架构和 ComfyUI 版本的 comfy-kitchen CUDA wheel。
• 包含 Native MiniMax-H3、LightX2V、Larryvrh、DMAD 和 LBH 示例工作流。
• 安装器会校验操作系统、GPU 架构、CUDA 版本和 wheel 平台，拒绝安装不匹配的包。
• 修复 0.1.0 的显存管理问题：allocation-heavy 的 reblock planner 不再进入私有 CUDA Graph，避免与 ComfyUI 的异步 CUDA allocator 产生不安全的 captured-allocation 生命周期；Spark 注意力 CUDA kernel 保持启用。
• 不包含 MiniMax-H3 模型或 LoRA 权重。

下载文件

GPU	下载文件
RTX 4090 / SM89	ComfyUI-Spark-H3-0.1.1-linux-x86_64-sm89-cu130.zip
RTX 50 系列、RTX PRO 6000 / SM120	ComfyUI-Spark-H3-0.1.1-linux-x86_64-sm120-cu130.zip

请根据 GPU 架构选择对应文件，不要混用 SM89 和 SM120 的安装包或后端 wheel。

安装方法

将下载的 ZIP 解压到 ComfyUI/custom_nodes，然后使用启动 ComfyUI 的同一个 Python 环境运行安装器：

cd /path/to/ComfyUI/custom_nodes
unzip /path/to/ComfyUI-Spark-H3-0.1.1-<对应版本>.zip
/path/to/ComfyUI/.venv/bin/python ComfyUI-Spark-H3/install.py

安装完成后重启 ComfyUI，并选择：

• RTX 4090：MiniMax H3 Spark Attention (SM89)
• RTX 50 系列或 RTX PRO 6000：MiniMax H3 Spark Attention (SM120)

性能结果

以下结果测量于 1344×768、345 帧、24 FPS、Top-K 20% 配置。

GPU	Fidelity 配置	Speed 全稀疏配置
RTX 4090 / SM89	2.37–2.61×	4.34–4.36×
RTX PRO 6000 / SM120	2.05–2.21×	3.14–3.15×

Fidelity 配置保留 Dense warmup；Speed 配置采用全稀疏执行，以吞吐速度为优先。实际性能会受到模型、工作流、稀疏参数和可用显存影响。

Windows 版本

Windows 版本正在开发和真实 GPU 验证中，将在安装、CUDA kernel 路径和性能验证完成后作为独立 Release 资产发布，敬请期待。

兼容性与限制

• 需要 NVIDIA SM89 或 SM120 GPU。
• 当前通过 GitHub Release ZIP 手动安装；ComfyUI Registry / Manager 发布将在后续提供。
• ComfyUI、MiniMax-H3 模型、LoRA 权重以及可选的第三方 LBH 节点均不包含在安装包内。

SHA256

6144173a2904ceb50c36495d4c889ccedea62ea5635adb04f2474e0e47d0401d  ComfyUI-Spark-H3-0.1.1-linux-x86_64-sm89-cu130.zip
3f32c0424b5c22dc308b263ef36faac9dc8f9266123cbd12d045196725df2ec4  ComfyUI-Spark-H3-0.1.1-linux-x86_64-sm120-cu130.zip

详细安装、模型放置和工作流说明请参阅 COMFYUI.md 与 中文安装指南。

【BUG 版本，请勿下载】ComfyUI Spark-H3 v0.1.0

⚠️ BUG 版本，请勿下载或继续使用

Caution

ComfyUI Spark-H3 0.1.0 存在已确认的显存管理缺陷，不应下载或继续使用。请升级到 ComfyUI Spark-H3 0.1.1。

问题说明

0.1.0 发布包使用了旧版 reblock 显存管理路径：allocation-heavy 的 reblock planner 会进入私有 PyTorch CUDA Graph。ComfyUI 0.38/0.39 默认使用异步 CUDA allocator，这会使 captured 临时显存的生命周期与 allocator 的释放、复用时机不一致，在重复运行或较大工作流中可能造成显存异常、进程中止或 OOM。

0.1.1 已将 reblock planner 移出私有 CUDA Graph，改为 eager 执行，同时保留 Spark 注意力 CUDA kernel。请勿尝试通过继续使用 0.1.0 或替换其 wheel 来规避该问题，直接升级到 0.1.1。

历史版本信息

0.1.0 是首个面向 NVIDIA SM120（RTX 50 系）GPU 的独立 ComfyUI 发布包，现仅作为问题版本归档保留。

• 原发布包：ComfyUI-Spark-H3-0.1.0.zip
• 原 SHA256：0f91950051d61251d766eeb54d9bb6abaaa707b881e63e4ac2457aa42575ca8c
• 替代版本：ComfyUI Spark-H3 0.1.1

Spark-H3 blog update (October 8, 2026)

Updated Spark-H3 technical blog.

Highlights:

• Spark-H3-Lite density sweep and refreshed latency results
• Acc-LoRA, DMAD-LoRA, SelfLift-LBH, and Ref2VA integrations
• Updated VDN comparison
• Expanded BSA, Sol-Attn, Veda, and Spark-Attn discussion

The attached static-site bundle is verified by the GitHub Pages deployment workflow.

Spark-H3 blog VDN playback fix

Updates the Spark-H3 blog so VDN video scrubbing preserves the user's playback state.

Spark-H3 blog update (September 28, 2026)

Adds Spark-H3-30pct benchmark results, clarifies the reported denoiser step count, and refreshes the blog presentation.

Spark-H3 blog

Official September 23 publication snapshot of the Spark-H3 technical blog.

This release adds the revised Spark-Reblock and Spark-Reweight descriptions and ablations, the Dense/Sol-H3/Spark-H3 benchmark tables, and integration results for FastH3, the Larryvrh and LightX2V few-step LoRAs, and Video DeltaNet.

Protocol clarification: the main benchmark uses a 20-point MiniMax-H3 sigma schedule, corresponding to 19 denoiser evaluations. Its reported ATTN speedup is an attention-core metric. The Video DeltaNet comparison instead reports complete attention-module latency per model evaluation; the two ATTN metrics have different timing boundaries and should not be compared directly.

Spark-H3 blog

Benchmark table correction. The reported ATTN speedups are 2.19× for Sol-H3, 2.33× for Spark-H3-10pct, and 1.96× for Spark-H3-20pct.

These values are schedule-aggregated, operator-inclusive attention-core speedup ratios measured with the September 21 CUDA-event harness. They include the recipe’s required dense warmup evaluations, dense first layer, and sink-query work, but exclude QKV and output projections, the rest of the DiT, and the surrounding generation pipeline. They are not full-pipeline end-to-end speedups or isolated sparse-kernel speedups.

Spark-H3 blog

Historical static-site snapshot of the Spark-H3 blog, including 32 video previews and two interactive animations.

The benchmark table in this snapshot predates the September 21 corrections. Its reported “Attn speedup” values are isolated sparse-kernel-path ratios rather than schedule-aggregated attention-core speedups. Refer to blog-20260921 or a later snapshot for the revised metric definition and benchmark values.

Package SHA-256: fff0a0ac3d838eb81734eb4093862922fc65c53361080d57b5e60469b8e82a76.