MiniMax H3 · Wuxia High-Dynamic Fight Workflow (WORK-FISHER)

中文说明 · YouTube · Bilibili 哔哩哔哩

A two-pass ComfyUI workflow for MiniMax H3 (reference image → video with audio), tuned for fast, high-motion martial-arts fights.

Pass 1 drafts motion and composition at low resolution.
A learned latent upscaler enlarges it ×1.5.
Pass 2 refines detail at high resolution.
A pass-1 preview node shows the draft right after pass 1, so you can cancel early if the motion is wrong.

This repo contains only the workflow file and download links. Custom nodes and models come from their original sources.

Requirements

ComfyUI 0.36.0 or newer. Tested on 0.36.0, Python 3.12, PyTorch 2.10 + CUDA 13.0.
NVIDIA GPU. Tested on RTX 4070 Ti 12 GB + 32 GB RAM.
Python packages:
- sageattention: used by the KJNodes memory-efficient Sage patch.
- comfy-kitchen: used by T8's Sol-Attn. Usually already bundled with recent ComfyUI.

1. Custom nodes (6)

Easiest way: open the workflow, then use ComfyUI-Manager → Install Missing Custom Nodes. Or clone each repo into ComfyUI/custom_nodes/.

Custom node	What it does in this workflow	Tested version
comfyui-minimax-h3-audio-T8	Core H3 nodes: conditioning, sampler, two-pass plan, latent upscale, decode, semantic bridge, Sol-Attn	1.86.0 (`8b7408e`)
ComfyUI-KJNodes	Low-VRAM attention, chunked feed-forward, Sage patch, and all Set/Get nodes	1.5.0 (`e8e88f7`)
ComfyUI-VideoHelperSuite	Video Combine (pass-1 preview and final video)	1.7.9 (`115de7a`)
ComfyUI_Comfyroll_CustomNodes	CR Prompt Text	`d78b780`
ComfyUI-utils-nodes	Text concatenation (fixed English prefix + your prompt)	1.4.2
ComfyUI_LayerStyle	Reference image resize	2.0.42 (`557d882`)

Sol-Attn: use only the one built into T8. Do not install sol_attn_minimax_pr117 or ComfyUI-SolAttn_triton. They register the same node, override T8's version, and cause a sol_attn error on every step.

2. Models (12 files, 43.9 GB)

Put each file in the listed folder under ComfyUI/models/. ModelScope mirrors are much faster in mainland China.

Folder	File	Size	Download	ModelScope mirror
`diffusion_models`	`minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors`	19.53 GB	Hugging Face	–
`text_encoders`	`qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors`	14.61 GB	Hugging Face	ModelScope
`vae`	`minimax_h3_video_vae_fp16.safetensors`	4.85 GB	Hugging Face	ModelScope
`vae`	`minimax_h3_audio_vae_fp32.safetensors`	0.56 GB	Hugging Face	ModelScope
`latent_upscale_models`	`h3_upscaler_sharpness_2000steps_v0.1_fp32.safetensors` (rename, see below)	1.29 GB	Hugging Face	ModelScope
`semantic_bridge/t8_compat`	`BUNNY_H3_ActionLogic_Bridge_V1_T8_Compat.safetensors`	0.02 GB	Hugging Face	–
`loras`	`minimax_h3_fl2v_turbo_4step_v1.2_768p_comfyui_bf16.safetensors`	1.82 GB	Hugging Face	ModelScope
`loras`	`minimax_h3_turbo_v4_step600_ema_DasiwaREF2VAHybridV1_0_curveproj1025_compat_v001.safetensors`	0.74 GB	Hugging Face	–
`loras`	`H3_Combat_V2.safetensors`	0.14 GB	Hugging Face	–
`loras`	`GunFu.safetensors`	0.14 GB	Hugging Face	–
`loras`	`Motion_Repair.safetensors`	0.14 GB	Hugging Face	–
`loras`	`H3_speed_slider_1.1.safetensors`	0.01 GB	Hugging Face	–

Notes:

Rename the upscaler. The file is published as h3_upscaler_lms_v0.1_fp32.safetensors. Rename it to h3_upscaler_sharpness_2000steps_v0.1_fp32.safetensors, or re-select it in the upscale node.
Semantic bridge folder. Create ComfyUI/models/semantic_bridge/t8_compat/ yourself. It is T8's own folder and does not exist by default. Keep the t8_compat level.
Direct download links. In any Hugging Face link, replace /blob/ with /resolve/.
Verified files. Every link was checked by SHA-256 against the files this workflow was tested with.

3. How to use

Reference image: load your image in group 参考图1 (reference image 1). Groups 2–6 are bypassed; select a group and press Ctrl+B to enable it.
Prompt and duration: write your prompt in 简单提示词写入 (prompt). Set 时长 (duration) in seconds. The maximum is 15; start with about 5 s.
Queue. After pass 1, the purple 第一段预览 (pass-1 preview) node plays a low-res draft. If the motion is wrong, press cancel to skip the slow pass 2.
Output: finished videos go to ComfyUI/output.

Group names are in Chinese:

Group	Meaning
加载图像 / 参考图1–6	Load images / reference images 1–6
简单提示词写入	Prompt
时长	Duration
比例/尺寸	Aspect ratio / size
调动作快慢和幅度滑杆	Motion speed slider
低噪	Pass 1 (low-res draft)
放大	Upscale
高噪	Pass 2 (high-res refine)
第一段预览	Pass-1 preview
输出视频	Output video

4. Prompt template

打戏提示词模版_MiniMax15秒一镜到底_v1.4.md is a fight-prompt template for this workflow. It is written in Chinese.

What it does: turns one reference image into a 15-second, one-take fight prompt with 8 beats.
Built-in rules:
- an opponent who wins exchanges;
- character facing written for every beat;
- a state lock, so riding or giant-avatar states do not flicker;
- one big move chosen from 7 types;
- an anime-style finishing strike.
How to use: give the whole template to an LLM together with your reference image. It outputs only the prompt body. Paste that into 简单提示词写入 (prompt).

Credits and licenses

Custom nodes: T8mars, kijai, Kosinkadink, Suzie1, zhangp365, chflame163.
Models: by the authors linked above. MiniMax H3 is licensed under the MiniMax H3 Community License Agreement.
Every custom node and model follows its own license. Check its page before commercial use.
Workflow by WORK-FISHER.

中文说明

MiniMax H3 两段式工作流：参考图生成带声音的视频，专门调过快节奏、高动态的武打戏。

第一段：低分辨率打草稿，定动作和构图。
放大：学习型潜空间放大，放大 1.5 倍。
第二段：高分辨率精修细节。
第一段预览：第一段跑完就能看到草稿，动作不对可以直接取消，不用等最慢的第二段。

这里只放工作流文件和下载链接，插件和模型都从原始出处下载。

环境要求

ComfyUI 0.36.0 或更新。测试环境：0.36.0、Python 3.12、PyTorch 2.10 + CUDA 13.0。
NVIDIA 显卡。测试机：RTX 4070 Ti 12GB 显存 + 32GB 内存。
需要两个 Python 包：
- sageattention：KJNodes 的 Sage 省显存补丁要用。
- comfy-kitchen：T8 的 Sol 稀疏注意力要用，新版 ComfyUI 一般已经自带。

一、插件（6 个）

最简单的装法：打开工作流后，用 ComfyUI-Manager → 安装缺失节点。也可以把各个仓库分别克隆到 ComfyUI/custom_nodes/。插件地址和测试版本见上方英文部分的表格。

Sol 稀疏注意力只用 T8 自带的版本。 不要另装 sol_attn_minimax_pr117 或 ComfyUI-SolAttn_triton：它们会注册同名节点，把 T8 的版本覆盖掉，导致采样时每一步都报 sol_attn 错误。

二、模型（12 个，共 43.9GB）

下载地址见上方英文部分的表格，文件放到 ComfyUI/models/ 下对应的文件夹。国内优先用 ModelScope 镜像，速度快很多。

放大模型要改名：仓库里叫 h3_upscaler_lms_v0.1_fp32.safetensors，下完改成 h3_upscaler_sharpness_2000steps_v0.1_fp32.safetensors；或者在工作流的放大节点里重新选一下。
语义桥要自己建文件夹：在 models 里新建 semantic_bridge，再在里面新建 t8_compat，文件放进去。t8_compat 这一层不能省。
直接下载：把 Hugging Face 链接里的 /blob/ 换成 /resolve/。
文件都核对过：每个链接都按 SHA256 跟作者实际跑通的文件比对过。

三、使用

换参考图：在「参考图1」组里上传你的图。参考图 2–6 是旁路状态，要用哪个就选中那一组，按 Ctrl+B 启用。
填提示词和时长：在「简单提示词写入」里写提示词，在「时长」里填秒数，最长 15 秒。建议先用 5 秒左右跑通再加长。
点运行：第一段跑完后，紫色的「第一段预览」节点会先出一段低清视频。动作不对就点取消，省下第二段的时间。
拿成片：成片在 ComfyUI/output。

四、打戏提示词模版

打戏提示词模版_MiniMax15秒一镜到底_v1.4.md 是配合这个工作流用的打戏提示词模版。

能做什么：把一张参考图写成 15 秒一镜到底、8 拍节奏的打戏提示词。
模版里已经定好的规矩：
- 对手是强敌，有来有回；
- 每一拍都写明人物朝向；
- 状态锁定，骑马、法相不会一会儿有一会儿没有；
- 大招七选一；
- 用动漫终结技收尾。
用法：把整份模版连同参考图一起交给大模型（LLM），它只会输出提示词本体。把输出直接贴进工作流的「简单提示词写入」就行。

致谢与许可

插件作者：T8mars、kijai、Kosinkadink、Suzie1、zhangp365、chflame163。
模型作者：见上表链接。MiniMax H3 适用 MiniMax H3 Community License。
各插件、各模型都有各自的许可，商用前请到原页面确认。
工作流作者：WORK-FISHER。

Downloads last month: -

Work-Fisher
/

MiniMax-H3-Wuxia-Fight-Workflow