⚡ MINIMAX H3 // 4 STEPS TURBO VIDEO AIO
「WITH TURBO CUSTOM NODES」
Designed & Engineered by EKKIVOK
USE WITH THIS LORA :https://civarchive.com/models/2838221/minimax-h3-turbo-lora?modelVersionId=3203523
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
SYSTEM STATUS
VIDEO ENGINE ● ONLINE
AUDIO ENGINE ● ONLINE
4 STEPS TURBO ● ACTIVE
PROMPT AI ● OPTIONAL
RTX UPSCALER ● OPTIONAL
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━🚀 OVERVIEW
The MINIMAX H3 // 4 STEPS TURBO VIDEO AIO workflow has been designed to make MiniMax H3 video generation as fast, intuitive and efficient as possible while keeping the workflow clean and beginner-friendly.
Instead of exposing dozens of disconnected nodes, everything has been carefully organized into logical production stages, making the workflow easy to understand, easy to customize and suitable for both beginners and experienced ComfyUI users.
Whether your goal is to generate a cinematic video from a text prompt or create an animation guided by reference images, this workflow provides a streamlined production pipeline from start to finish.
⚡ MAIN FEATURES
✅ Text → Video
✅ First Frame → Video
✅ First + Last Frame animation
✅ Native Audio Generation
✅ 4 Steps Turbo Workflow
✅ Optional Prompt Enhancer (Qwen VLM)
✅ RTX Video Super Resolution (optional)
✅ Automatic Video Export
✅ Optimized node organization
✅ Cyberpunk / Anime themed interface
🎬 WHAT THIS WORKFLOW CAN DO
Generate high-quality videos in only 4 inference steps
Create videos directly from text prompts
Animate from a single reference image
Control both the first and last frame for better temporal consistency
Generate synchronized native audio
Enhance prompts automatically using Qwen VLM
Upscale the final result using NVIDIA RTX Video Super Resolution
Produce ready-to-share MP4 videos
📖 COMPLETE DOCUMENTATION INCLUDED
One of the goals of this project was to make the workflow self-contained.
The workflow itself includes:
📘 Complete step-by-step tutorial
⚙️ Installation instructions
📥 Model download information
💡 Usage tips
🎛️ Parameter explanations
⚡ Performance recommendations
❓ Troubleshooting notes
Every important piece of documentation is directly integrated inside the workflow through Markdown panels, allowing you to learn and use the workflow without constantly switching between Civitai and ComfyUI.
Simply open the workflow and follow the built-in documentation.
🎮 WORKFLOW MODES
TEXT TO VIDEO
Generate a complete video from a prompt.
FIRST FRAME MODE
Guide the animation using a starting image.
Perfect for maintaining character identity and composition.
FIRST + LAST FRAME MODE
Control both the beginning and the ending of the animation for improved consistency.
PROMPT ENHANCER
Enable it if you want AI to automatically rewrite and improve your prompt.
Disable it if you prefer to keep your original prompt untouched.
RTX VIDEO UPSCALER
Optional AI upscaling after generation for maximum image quality.
Recommended for final renders only.
💻 RECOMMENDED HARDWARE
✅ Minimum Configuration
NVIDIA RTX GPU
12 GB VRAM
CUDA compatible
32 GB System RAM recommended
Ideal for:
720p generation
Short videos
Testing prompts
Learning the workflow
🚀 Recommended Configuration
NVIDIA RTX GPU
24–32 GB VRAM
CUDA compatible
Fast NVMe SSD
Modern multi-core CPU
Recommended for:
1080p generation
Longer videos
Multiple LoRAs
RTX Video Super Resolution
Professional workflows
🔥 Best Experience
RTX 5090 32 GB (or equivalent)
High-end CPU
64 GB+ System RAM
PCIe 4.0 / 5.0 NVMe SSD
Perfect for long productions, maximum generation speed and heavy workloads.
❤️ WHY THIS WORKFLOW?
This is not just another collection of nodes.
Every section has been carefully designed to provide a clean production pipeline while keeping advanced features easily accessible.
The objective is simple:
Faster workflow
Better organization
Easier learning
Professional presentation
Maximum productivity
⚠ IMPORTANT
Please keep the original group names if you modify the workflow.
Some bypass controls rely on these group names to function correctly.
❤️ CREDITS
Workflow Design
EKKIVOK
Technologies Used
MiniMax H3
ComfyUI
Qwen VLM
NVIDIA RTX Video Super Resolution
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
MADE WITH ❤️ BY EKKIVOK
CREATE • EXPERIMENT • INNOVATE
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━Description
THIS IS THE FINAL WORKFLOW.
FAQ
Comments (19)
这个RH的专用节点吧？RHMiniMaxH3SageAttentionPatch，直接从RH下载的工作流就分享了？
@EKKIVOK 普通sage attention 不能平替吗？
@BIG_A well for sage you need to install it, but with 4 steps lora you will not need it, just select "Off"
I uploaded My Nodes version hope it will help it's on this page.
DO know what this error is？:RuntimeError: The size of tensor a (236208) must match the size of tensor b (5689472) at non-singleton dimension 2.
for the sage attention patch even after downloading the github and requirements from https://github.com/HM-RunningHub/ComfyUI_RH_MinMaxH3 it still says the node is UNKNOWN and does not work.
it also gives a noattritbute qk_int8_sv_f16_accum_f32_attn in case you know what I need in order for that to work, I use sage attention 2.2 as it is
i have the same version and it's working ...but i don't use it, because with the lightining lora you don't really need that. i use sage only for regular 20 steps. but i think i will upload my version of the node that is working with the workflow, i will upload it in zip format :)
It's uploaded :) Have Fun !
I'm having trouble trying to fix the following unknown node message for RHMiniMaxH3SageAttentionPatch
i have uploaded The nodes :)
Thanks, I didn't notice earlier :-)
@EKKIVOK no still got the same Missing Node Pack : RHMiniMaxH3SageAttentionPatch; I reinstalled the comfyui-minimax-h3-turbo node nothing happend same error; But you guys can delete that node for the "sageAttention" and then tie the noodles back together so it works.
how to install Qwen3-VL-4B-Instruct-c_abliterated-v2
Looks like the url is wrong when trying to reach the server to download the model, try a different one. The creator of the node must fix it.
Duuude... This is the best workflow i've ever tested so far. Wooow. best speed best quality. h3 is so much fun with it. thanks a lot. just sad it's the final version. would be awesome to also have a multi image and multi reference video pipeline it. than this would be the ultimate 100% AIO workflow. :-)
HIGH-SPEED PREVIEW don't work. please help me.
Not sure if that's the problem, but the VAE file needs to be inside the ComfyUI\models\vae_approx folder and not ComfyUI\models\vae.
Took me a while to find that out.
@BonerSoup A complete reinstallation of the ComfyUI-KJNodes package helped.