How to Run MiniMax H3 Locally via ComfyUI (Step-by-Step Setup Guide)

MiniMax H3 is an open-weight AI video generation model that runs locally through ComfyUI, allowing you to generate high-quality videos with synchronized audio without relying on cloud APIs. Whether you’re an AI creator exploring open-weight models, a Stable Diffusion user expanding into video generation, or a developer building custom AI workflows, this guide walks you through the complete setup process, from installation and system requirements to workflow configuration and your first video generation.

By the end of this guide, you’ll have a fully functional MiniMax H3 installation in ComfyUI and understand how to generate text-to-video, image-to-video, and reference-based videos with native audio support on your own hardware.

Quick Summary

Difficulty Level: Intermediate

Supported Platforms: Windows and Linux (macOS support is limited)

Minimum GPU: NVIDIA RTX 3060 (12 GB VRAM)

Recommended GPU: NVIDIA RTX 4080 (16 GB VRAM) or higher

Best For: Local AI video generation, content creation, character animation, product demos, and creative projects

Supported Workflows: Text-to-video, image-to-video, reference-to-video, first-frame conditioning, and last-frame conditioning

Key Features: Native audio generation, offline inference, open-weight model, and ComfyUI workflow support

What Is MiniMax H3?

MiniMax H3 is an open-weight video generation model developed by the MiniMax team. It’s designed specifically for local inference, meaning you download the model weights and run everything on your hardware.

Key capabilities include:

Native Audio Generation. Unlike many video models, MiniMax H3 generates audio that’s synchronized to your video. No need for separate audio workflows or external tools. The audio VAE handles voice, music, and sound effects directly from your prompt.

Multiple Workflow Types. Text-to-video transforms text prompts into video. Image-to-video uses a starting image to condition the generation. Reference-to-video lets you provide a reference video to guide style and composition. First-frame and last-frame workflows provide additional control.

2K Output Support. You can generate videos at 1024×1024 or higher resolutions, balancing quality against rendering time and VRAM usage.

Practical Model Variants. MiniMax H3 offers INT8 and FP8 quantized versions that significantly reduce VRAM requirements while maintaining quality. Full-precision versions are available for highest quality output.

Open-Weight Design. The model weights are publicly available on Hugging Face. You own the model, control the inference, and don’t depend on API rate limits or commercial services.

Why Use ComfyUI?

ComfyUI has become one of the most popular interfaces for running open-weight image and video generation models. flexibility, visual clarity, and active community support.

Official Workflow Support. If official ComfyUI workflows are available from the MiniMax team, they are generally the best starting point because they are designed for compatibility with the latest model releases.

Visual Node Editor. You don’t write code to generate videos. ComfyUI’s node-based interface lets you drag connections, adjust parameters, and experiment instantly. This visual approach makes complex workflows accessible while remaining powerful enough for advanced customization.

Custom Workflows. You can modify official workflows, save your own presets, and build specialized workflows for your use cases. Save a workflow that works, reuse it, and share it with collaborators.

Easy Experimentation. Testing different prompts, settings, and workflow configurations takes seconds. Lower resolution previews let you validate concepts quickly before committing to high-quality renders.

Community Templates. The ComfyUI community has published hundreds of workflows. You can learn from existing examples and adapt them for MiniMax H3.

Cross-Platform. ComfyUI runs on Windows, Linux, and macOS, letting you use consistent workflows across different hardware.

System Requirements

To run MiniMax H3 locally, your hardware needs to handle large language models and video encoding simultaneously. Here’s what you need:

RTX 3060 Setup. 12GB VRAM barely handles MiniMax H3. Expect to use INT8 quantization, lower resolutions (512×512 or 768×768), and fewer inference steps. Render times stretch to 15-20 minutes per video.

RTX 4070 Setup. 12GB VRAM with slightly better compute. Similar constraints to RTX 3060, but faster processing. Reasonable for testing and iterative work.

RTX 4080 Setup. 16GB VRAM allows full-precision models, 1024×1024 output, and reasonable inference steps. Fast enough for production work without extreme optimization sacrifices.

RTX 4090 Setup. 24GB VRAM handles everything. Full-precision models, high resolution, many inference steps, and batch processing. Render times drop to 5-8 minutes for 2K videos.

RTX 5090 Setup.RTX 5090 provides additional VRAM and compute performance, allowing higher resolutions and longer videos to be generated more comfortably than previous-generation GPUs. Actual render times vary depending on workflow settings.

Storage is often overlooked. MiniMax H3 models alone consume 50-60GB. ComfyUI, dependencies, and workflow variations add another 30-40GB. Plan for 150GB free space if you’re experimenting with multiple models or maintaining several projects.

Files Required Before Installation

Before installing MiniMax H3, make sure you have the required model files and supporting components referenced in the latest official documentation:

MiniMax H3 Model Checkpoint. The main model weights used for video generation. Choose the variant that matches your available GPU memory.

Text and Vision Encoders. Required for processing text prompts and visual inputs. The exact files depend on the MiniMax H3 release and workflow.

Video VAE. Required for encoding and decoding video during generation.

Audio VAE (if required). Some workflows use a separate Audio VAE for synchronized audio generation.

Official ComfyUI Workflow. The JSON workflow that defines the node setup for MiniMax H3.

ComfyUI Framework. The interface and execution engine used to run the model locally.

Before continuing, ensure all model files and workflow assets belong to the same MiniMax H3 release to avoid compatibility issues.

Step-by-Step Installation

Step 1: Install ComfyUI

Start with a clean Python environment. Open your terminal or command prompt.

Begin by installing the latest version of ComfyUI in a clean Python environment.

git clone https://github.com/comfyanonymous/ComfyUI.git

cd ComfyUI

python -m venv venv

On Windows, activate the virtual environment:

venv\Scripts\activate

On Linux or macOS:

source venv/bin/activate

Install PyTorch with CUDA support:

pip install torch torchvision torchaudio –index-url https://download.pytorch.org/whl/cu121

Install ComfyUI requirements:

pip install -r requirements.txt

This takes 10-15 minutes depending on your internet speed and hardware.

Step 2: Update ComfyUI

Even if you just cloned, updates may be available:

cd ComfyUI

git pull origin main

pip install –upgrade -r requirements.txt

Check for custom node manager compatibility. ComfyUI Manager (optional but recommended) simplifies model and custom node installation:

cd custom_nodes

git clone https://github.com/ltdrdata/ComfyUI-Manager

Step 3: Download the MiniMax H3 Models

Next, download the required MiniMax H3 model files and organize them within your ComfyUI installation. Before downloading, create the following directory structure if it doesn’t already exist:

ComfyUI/

└── models/

    ├── checkpoints/

    ├── vae/

    └── encoders/

Download the latest MiniMax H3 checkpoint from the official MiniMax Hugging Face repository or the download source referenced in the official documentation. Because the model files are large, Git LFS (Large File Storage) is often required.

Install Git LFS if you haven’t already:

git lfs install

After downloading the checkpoint, copy it into the following directory:

ComfyUI/models/checkpoints/

Keep the original filename provided with the download unless the official documentation specifically instructs you to rename it. If additional model files, such as quantized variants, are available, choose the version that best matches your available GPU memory and performance requirements.

Step 4: Install the Required Encoders

MiniMax H3 requires the appropriate text and vision encoders specified by the official model release or ComfyUI workflow. Download the required encoder files from the official MiniMax H3 documentation or the linked model repository, then place them in the appropriate ComfyUI model directories (such as ComfyUI/models/text_encoders/ or the directory specified by the workflow).

The exact encoder files and installation paths may vary between MiniMax H3 releases, so always follow the latest official setup instructions to ensure compatibility with your workflow.

Step 5: Install the Required VAE Files

Download the Video VAE and Audio VAE files referenced in the official MiniMax H3 documentation or the recommended ComfyUI workflow. These components are required for encoding, decoding, and generating synchronized video and audio.

Copy the downloaded VAE files into the appropriate ComfyUI model directory, typically:

ComfyUI/models/vae/

If the official workflow specifies a different location for the audio VAE or other supporting files, follow those instructions. Because model assets and directory structures may change between releases, always use the latest official documentation to verify the required files and installation paths.

Step 6: Import the Official Workflow

Download the official MiniMax H3 ComfyUI workflow included with the latest MiniMax H3 release, then save the workflow JSON file to a convenient location on your computer.

Start ComfyUI:

python main.py

Then:

1. Open the ComfyUI web interface in your browser.
2. Click Load and select the downloaded workflow JSON file.
3. Wait for the workflow to load.

If you encounter missing nodes or model errors, verify that you’ve installed all required custom nodes and placed the required model files in the correct ComfyUI directories. Always use the workflow and model assets from the same MiniMax H3 release to ensure compatibility.

Step 7: Configure Workflow

The workflow contains several key nodes you’ll adjust:

Prompt Node: Enter your video description. “A cinematic shot of a person walking through a futuristic city at sunset” is more effective than generic prompts.

Sampler Node: Controls generation quality. DPM++ Karras is a common choice. More steps (25-50) produce better quality but take longer.

Seed Node: Determines randomness. Use -1 for random, or set a specific number to reproduce results.

Resolution Settings: 512×512, 768×768, or 1024×1024. Higher resolution increases render time exponentially.

FPS (Frames Per Second): 24 or 30 FPS. Higher FPS creates smoother motion but longer generation time.

Guidance Scale: 7.5-15 range. Higher values follow prompts more closely; lower values allow more creativity.

Duration: Usually 4-8 seconds. Longer videos increase VRAM usage and render time.

Step 8: Generate Your First Video

For your first test, use a low resolution (such as 512×512) and a short duration to reduce rendering time and VRAM usage.

Enter a simple prompt, then click Queue to start generation. ComfyUI will display the rendering progress and GPU memory usage during the process.

Once the render is complete, your generated video will be saved in the ComfyUI/output/ directory. Review the output to ensure the video and audio have been generated correctly.

Understanding MiniMax H3 Workflows

Text-to-Video Workflow

Text-to-video is the most intuitive workflow. You write a descriptive prompt, and MiniMax generates video matching your description.

Best Use Cases: Creating original content concepts, exploring ideas, generating background footage, producing cinematic shots, designing visual effects, animating objects or characters.

Advantages: Fastest to iterate, no reference material needed, most control over output, works with abstract concepts.

Example Prompt: “Cinematic overhead shot of a busy Tokyo street intersection at night. Neon signs reflect on wet pavement. Crowds of people walk in organized chaos. Camera slowly pans across the scene. Native audio captures ambient street sounds and distant music.”

Why this works: Specific details (overhead, Tokyo, night, neon, wet pavement) guide the model. Mentioning camera movement and audio expectations improves results.

Image-to-Video Workflow

Image-to-video takes a static image and extends it into video, maintaining the composition and style.

Best Practices: Use high-quality reference images. MiniMax H3 respects the input image structure while adding motion and detail. Start with 4-6 second durations. Adjust guidance to balance fidelity to the source image against natural motion generation.

Recommended Settings: Guidance 7.5-10 (lower than text-to-video), 25-35 inference steps, 768×768 to match input image quality.

Reference-to-Video Workflow

Reference-to-video uses an existing video as style and motion reference while generating new content with your text prompt.

Character Consistency: If you want a character to maintain consistent appearance across multiple videos, provide a reference video of that character. MiniMax learns the style from the reference and applies it to new prompts.

Style Preservation: Landscape photography styles, cinematic looks, color grading, and composition patterns transfer from reference videos to new generations.

When to Use: Creating series with consistent visual style, adapting existing footage aesthetically, maintaining character identity across videos, ensuring color consistency in multi-video projects.

Recommended Settings

Different workflows and hardware require different parameter combinations:

Fast Testing: Quick concept validation. Expect lower visual fidelity but rapid iteration. Essential for prompt refinement.

High Quality: Polished final output. Highest setting recommendations. Reserve for renders you plan to publish.

Low VRAM: For 12GB hardware or when running other applications. Accept quality reduction in exchange for functionality.

Cinematic: Premium workflows. Best for professional or portfolio-worthy content. Longest render times.

Character Animation: Balanced settings for character-focused videos. Moderate VRAM, reasonable render time, good quality.

Performance Expectations

Rendering performance depends on your GPU, VRAM, workflow configuration, video resolution, clip duration, and sampling settings. Higher resolutions and longer videos require more processing time and GPU memory.

RTX 3060 / RTX 4070 (12GB): Best suited for lower resolutions, shorter clips, and quantized model variants.

RTX 4080 (16GB): A solid choice for high-quality local video generation with balanced performance and VRAM capacity.

RTX 4090 / RTX 5090 (24GB+): Ideal for higher resolutions, longer videos, and more demanding workflows.

Actual render times vary based on your hardware and generation settings. If you encounter GPU memory limitations, reduce the resolution, shorten the video duration, lower the sampling steps, or use a lower-memory model variant.

VRAM Optimization Tips

Limited VRAM doesn’t prevent MiniMax H3 use;it requires strategy.

INT8 Quantization. MiniMax offers INT8 quantized model versions that use roughly 60% of full-precision VRAM while maintaining 95% of quality. Download the INT8 checkpoint instead of full precision.

SageAttention. A memory-efficient attention implementation. Some ComfyUI custom nodes support SageAttention. Enable it in configuration to reduce VRAM by 20-30%.

Lower Resolution. 512×512 uses roughly 40% of the VRAM of 1024×1024. If you only need web-quality video, lower resolution saves significant resources.

Fewer Inference Steps. Each step consumes additional memory. Reducing from 40 to 25 steps frees 10-15% VRAM. Trade some quality for feasibility.

Memory Offloading. ComfyUI can offload models to system RAM or disk. Slower but functional. Enable in settings if hitting hard VRAM limits.

Batch Size Optimization. If generating multiple videos, process them sequentially rather than batching. Single-video queues simplify VRAM tracking.

Monitor VRAM Before Render. ComfyUI displays estimated VRAM requirements. If it exceeds your capacity, adjust settings before queuing.

Read More: 10 Best AI IDEs for Developers in 2026 

Common Errors & Fixes

CUDA Out of Memory: This is the most common issue. It means your GPU doesn’t have enough memory for the current settings. In ComfyUI, reduce resolution first (biggest VRAM impact). If still failing, reduce inference steps.

Model File Not Found: ComfyUI looks in specific directories. Verify that model filenames match exactly what the workflow expects. Case sensitivity matters on Linux.

Workflow Loading Errors: Official workflows sometimes require specific custom node versions. Run git pull in the ComfyUI directory to update, and check the MiniMax repository for custom node installation instructions.

MiniMax H3 Prompting Tips

Video prompting differs from image prompting. Success requires specificity and clarity.

Cinematic Scene: “Wide establishing shot of a luxury hotel lobby with marble columns, crystal chandeliers, and soft golden lighting. A concierge moves behind the desk. Camera slowly tracks from left to right. Subtle classical music plays in background.”

Why it works: Specific architecture details, lighting direction, camera movement, audio suggestion.

Realistic Character Video: “Medium close-up of a woman with shoulder-length brown hair, wearing a professional blazer. She sits at a wooden desk, looks at the camera, smiles warmly, then turns her head to look at a document on her desk. Soft natural window lighting from the left. Office sounds and a subtle pen click in background.”

Why it works: Physical details, framing (medium close-up), expression, gestures, lighting source, audio elements.

Product Shot: “Product photography of a sleek silver smartwatch on a minimalist white surface. Ring light photography setup with sharp shadows. Camera slowly rotates 360 degrees around the watch. Reflection visible on the polished surface. No audio.”

Why it works: Product context, lighting setup, camera movement, environmental details, explicit audio instruction (no audio).

Nature Scene: “Timelapse of sunrise over a misty mountain valley. Golden sunlight breaks through clouds. A river winds through the valley. Birds fly in formation across the sky. Natural sound: wind, bird calls, water flowing.”

Why it works: Temporal progression (timelapse), multiple elements (clouds, river, birds), specific sounds, cinematic setup.

Character Animation: “Cartoon-style robot character with blue metallic body and orange accents. The robot walks confidently from left to right across a futuristic lab. Its arms swing naturally. LED eyes blink. Mechanical whirring and beeping sounds in background.”

Why it works: Art style defined (cartoon-style), character appearance specific, movement described clearly, audio expectations clear.

MiniMax H3 vs Other Local Video Models

Choosing between local video models depends on your priorities:

MiniMax H3 Strengths: Native audio generation is a major advantage. No separate audio workflow. Output quality is competitive with commercial models. Active development and community.

MiniMax H3 Considerations: Larger model files. Higher VRAM requirements than some alternatives. Steeper learning curve for workflow customization.

Best Practices

Following a few simple best practices can make your MiniMax H3 ComfyUI workflow more efficient and reliable.

Start with the Official Workflow: Use the official MiniMax workflow before experimenting with custom configurations.

Test at Lower Resolutions First: Generate short, low-resolution previews before rendering high-quality videos to save time and VRAM.

Keep Models Organized: Store checkpoints, VAEs, and encoders in their recommended directories to avoid loading errors.

Save Working Workflows: Save successful workflow configurations so you can reuse them in future projects.

Monitor VRAM Usage: Watch GPU memory usage during rendering and reduce resolution or sampling steps if you encounter memory issues.

Update Carefully: Keep ComfyUI, models, and custom nodes up to date while backing up working workflows before major updates.

Save Successful Prompts: Build a library of prompts and settings that consistently produce good results for faster future iterations.

These practices help you maintain a stable setup, reduce troubleshooting, and generate more consistent results with MiniMax H3.

Limitations

While MiniMax H3 offers powerful local video generation capabilities, it’s important to understand its current limitations before setting up your workflow.

High Hardware Requirements: MiniMax H3 performs best on modern NVIDIA GPUs with sufficient VRAM. Lower-end GPUs can still run the model using quantized versions or reduced settings, but expect longer render times and lower output resolutions.

Time-Consuming Renders: Video generation is computationally intensive. Render times vary depending on your GPU, video resolution, clip duration, sampling steps, and workflow configuration, so generating longer or higher-quality videos may require considerable processing time.

Large Downloads and Storage Needs: The model checkpoints, encoders, VAEs, and other required assets occupy a significant amount of disk space. In addition, generated videos and cached files can quickly increase storage usage over time.

CUDA-Focused Ecosystem: Most MiniMax H3 workflows are optimized for NVIDIA CUDA. Although alternative backends may exist, compatibility and performance on AMD or Intel hardware can vary and may require additional configuration.

Steeper Learning Curve: ComfyUI is highly flexible, but new users should expect to spend time learning node-based workflows, prompt engineering, model management, and performance optimization before achieving consistent results.

Dependency Management: Running MiniMax H3 locally requires maintaining Python packages, ComfyUI updates, custom nodes, and compatible model versions. Keeping these components in sync is important for avoiding workflow errors.

Platform Support: Windows and Linux currently offer the most mature experience for local deployment. Support on macOS may be more limited depending on the available hardware and software compatibility.

Conclusion

Setting up MiniMax H3 in ComfyUI gives you a flexible way to generate AI videos locally while maintaining full control over your workflows, model files, and output quality. Although the initial installation requires downloading large models and configuring the necessary dependencies, the payoff is a customizable environment that doesn’t rely on cloud APIs or recurring usage limits.

Whether you’re experimenting with text-to-video, image-to-video, or reference-to-video generation, MiniMax H3 provides a solid foundation for local AI video creation with native audio support. As you become more familiar with ComfyUI, you can fine-tune prompts, sampling settings, and workflows to achieve more consistent and higher-quality results.

If you’re new to the model, start with the official workflow, lower resolutions, and shorter video durations to verify that everything is working correctly. Once your setup is stable, gradually increase the resolution, inference steps, and clip length based on your available GPU memory and performance requirements.

As the ecosystem continues to evolve, keeping your ComfyUI installation, workflows, and model files up to date will help ensure compatibility with future MiniMax H3 releases and improvements. With the right hardware and a bit of experimentation, MiniMax H3 ComfyUI can become a powerful local video generation setup for creators, developers, and AI enthusiasts looking for greater control over their projects.

Frequently Asked Questions

Can MiniMax H3 run on 12GB VRAM?

Yes, with constraints. Use INT8 quantization, limit resolution to 512×512, and set inference steps to 20-25. Render times stretch to 15-20 minutes per video. This is feasible for experimentation but not production workflows.

Does MiniMax H3 generate audio?

Yes, natively. The audio VAE generates synchronized sound effects, music, and voice based on your prompt. No separate audio tools required.

Can I use AMD GPUs?

Limited support. AMD ROCm is theoretically compatible, but few users report successful setups. NVIDIA GPUs remain the reliable choice.

Is MiniMax H3 free?

Yes. The model is open-weight and available on Hugging Face. No subscription, API costs, or licensing fees. You only pay for electricity to run it.

Does it work on Linux?

Yes, fully supported. Installation process is identical. Linux often has better performance than Windows due to driver overhead differences.

Where can I download MiniMax H3?

Download the latest MiniMax H3 model files from the official MiniMax Hugging Face repository. Since the files are large, it’s recommended to use Git LFS for a reliable download. Place the downloaded files in the appropriate ComfyUI/models/ directories before loading the workflow.

Which workflow should beginners use?

Start with text-to-video at 512×512 resolution. It’s the most intuitive and fastest to iterate. Once comfortable, explore image-to-video and higher resolutions.

How much storage is required?

Budget 150GB minimum. 50-60GB for models, 30-40GB for ComfyUI and dependencies, 60GB+ for output videos if you generate frequently.

Can I use custom workflows?

Yes. Once you understand the official workflows, you can create custom nodes, modify existing workflows, and save your variations. The ComfyUI ecosystem supports extensive customization.

Why is my workflow failing?

Check three things: Are all model files present and in the correct directories? Are custom nodes installed? Is your VRAM sufficient? 95% of failures trace to missing files or insufficient memory.