- MiniMax H3 in ComfyUI: Run open-weights AI video generation locally with full parameter control
- Three core modes: Text-to-Video (T2V), Image-to-Video (I2V), and Reference-to-Video (R2V)
- Native stereo audio: Dialogue, SFX, and music are generated together with the video
- Hardware requirement: NVIDIA RTX 3060 or higher with sufficient system RAM
- Speed optimization: Install Sage Attention to roughly double generation speed
MiniMax H3 ComfyUI Overview
MiniMax H3 brings powerful open-weights AI video generation directly to your local machine through ComfyUI. This integration allows creators to generate high-quality video content with native stereo audio, combining text prompts, reference images, and video clips into cohesive multimodal outputs. Running locally provides full control over every parameter without relying on cloud APIs.
Video Highlights:
- Local text-to-video generation using the pruned int8 model on an RTX 3090
- First and last frame image-to-video transitions with wipe and phase effects
- Reference-to-video combining multiple images, video clips, and audio tracks
- Voice cloning capabilities by merging character appearances and voices from separate clips
- Pushing beyond the 15-second maximum duration limit successfully
The MiniMax H3 license restricts downloads in the USA and European Union by default. However, individuals can request access by submitting a license request form. Approval emails are typically sent quickly. Always conduct your own research regarding your specific usage rights.
Core Generation Modes
MiniMax H3 operates through three primary generation workflows, each designed for different creative use cases. Understanding when to use each mode is essential for achieving optimal results.
Text-to-Video (T2V)
- No input media required
- Generates from text prompts only
- Uses the FL2VA model
- Best for rapid concept testing
Image-to-Video (I2V)
- First and/or last frame control
- Optional keyframe inputs
- Uses the FL2VA model
- Ideal for animating static images
Reference-to-Video (R2V)
- Up to 9 images, 3 videos, 3 audio clips
- Locks character identity and style
- Uses the Ref2VA model
- Best for consistent character work
Mode Comparison Table
| Feature | Text-to-Video | Image-to-Video | Reference-to-Video |
|---|---|---|---|
| Model Weights | FL2VA | FL2VA | Ref2VA |
| Max Reference Images | 0 | 2 (first/last) | 9 |
| Max Reference Videos | 0 | 0 | 3 |
| Standalone Audio Inputs | 0 | 0 | 3 |
| Native Audio Output | Yes | Yes | Yes |
| Best Use Case | Concept exploration | Keyframe animation | Character consistency |
T2V and I2V workflows share the same FL2VA model weights. R2V uses a separate Ref2VA diffusion model. When building custom workflows, use a boolean switch connected to your model loader groups to easily toggle between them.
Installation and Setup
Getting MiniMax H3 running in ComfyUI requires updating to a compatible version and downloading the appropriate model weights from Hugging Face.
Update ComfyUI
Update your ComfyUI installation to version 0.30.0 or later. Earlier versions lack native support for MiniMax H3 workflows and nodes.
Access Template Library
Navigate to Template Library > Video within the ComfyUI interface. Select any MiniMax H3 workflow that matches your intended generation mode.
Download Models
Follow the pop-up prompts to automatically download the required model files. Models are hosted on Hugging Face in the Comfy-Org/MiniMax-H3 repository.
Configure Resolution
Use the Resolution Selector node to set your output dimensions. Keep the multiple at 32 to match the H3 resolution grid. Choose from presets like 16:9, 9:16, or 1:1.
Run Generation
Connect your inputs, write your prompt, and execute the workflow. Generation time depends on your hardware, resolution, and number of references.
Hardware Recommendations
| GPU Tier | Minimum | Recommended | High-End |
|---|---|---|---|
| GPU | RTX 3060 | RTX 4070 | RTX 3090/4090 |
| Model Version | Pruned int8 | Standard | Standard |
| Est. Gen Time (10s) | 8-12 min | 5-7 min | 3-5 min |
| Max Practical Res | 864x480 | 1080p | 1216x672+ |
For users with 8GB VRAM cards like the RTX 3060, the pruned int8 model provides the best balance of quality and performance. Those with 24GB cards (RTX 3090/4090) can run full-precision models at higher resolutions without compromise.
Prompt Writing Strategies
Effective prompting in MiniMax H3 requires a structured approach that describes scenes, camera movements, and audio cues within a single prompt block. The model handles complex, multi-sentence prompts effectively.
Prompt Structure by Mode
| Mode | Prompt Structure | Key Syntax |
|---|---|---|
| T2V | Scene description, shot breakdown, camera moves, audio cues | Plain text |
| I2V | Motion description, transition style, audio elements | Plain text |
| First/Last Frame | Describe transition from first to last image | <Picture 0> and <Picture 1> tags |
| R2V | Assign roles to each reference explicitly | <Picture 1>, <Video 1>, <Audio 1> tags |
Prompting Best Practices
- Describe the whole scene first: State the location, character, and action before breaking into timed shots
- Include audio instructions: Specify dialogue, sound effects, and music within the same prompt block
- Use reference tags in R2V: Reference each input by tag in the exact connection order (e.g.,
<Picture 1>,<Video 1>) - Assign each reference a job: Explicitly state which reference drives identity, style, motion, camera, or voice
- Leverage first/last frame tags: Even though the model auto-maps start and end frames, using
<Picture 0>and<Picture 1>builds good habits for R2V workflows
The duration input snaps to the model's 17-frame-per-block (17k+5) grid at 24fps. While the stated maximum is 15 seconds, testing shows that pushing to 20 seconds can produce coherent results without garbled output.
Speed Optimization with Sage Attention
Generation speed is a critical factor when running large video models locally. Sage Attention provides a significant performance boost with minimal quality degradation.
Install SageAttention
Download the Python wheel matching your PyTorch and CUDA versions from the SageAttention releases page. Install using pip install <wheel-file>.
Install KJNodes
Use ComfyUI Manager to install the KJNodes custom nodes package. Alternatively, clone the repository into your ComfyUI/custom_nodes/ directory and restart ComfyUI.
Add the Patch Node
Insert a Patch Sage Attention KJ node between the UNETLoader and BasicGuider node. Set the sage_attention parameter to auto.
Run the Workflow
Execute your workflow as usual. Only the guider needs the patch; the scheduler remains unchanged.
Sage Attention requires float16 or bfloat16 tensors. MiniMax H3 runs some layers in other dtypes, so you may see "Input tensors must be in dtype of torch.float16 or torch.bfloat16, using pytorch attention instead" messages. This is expected behavior and generation will still work correctly.
Optimization Comparison
| Setting | Standard Attention | Sage Attention | Global Flag |
|---|---|---|---|
| Speed | Baseline | ~2x faster | ~2x faster |
| Quality | Full fidelity | Minimal loss | Minimal loss |
| Setup | None | Node-based | --use-sage-attention |
| Console Warnings | None | Expected fallback messages | Expected fallback messages |
Workflow Architecture and Checklist
Understanding the node architecture helps you build and troubleshoot custom MiniMax H3 workflows. The workflow consists of model loaders, conditioning nodes, a standard K sampler, and output configuration.
Key Node Components
| Component | Function | Notes |
|---|---|---|
| Model Loaders | Load FL2VA or Ref2VA weights | Separate groups for first/last vs reference |
| CLIP and VAE | Text encoding and decoding | Shared across both model types |
| RG3 Fast Group Bypass | Toggles active model group | Ensures only one group runs at a time |
| Resolution Selector | Computes width and height | Outputs connect to H3 node inputs |
| K Sampler | Standard sampling | No special configuration needed |
| Video Output | Renders final MP4 | Supports GIF, frame interpolation |
While default workflows show a limited number of inputs, the system supports many more. You can easily expand to 5-6 images or add additional video and audio references by duplicating input nodes and updating your prompt tags accordingly.
MiniMax H3 ComfyUI Setup Checklist:
- Update ComfyUI to version 0.30.0 or later
- Download model weights from Hugging Face (Comfy-Org/MiniMax-H3)
- Verify GPU has at least 8GB VRAM (RTX 3060+)
- Ensure sufficient system RAM for model loading
- Select appropriate model variant (pruned int8 for lower VRAM)
- Set resolution multiple to 32 in Resolution Selector
- Install Sage Attention and KJNodes for faster generation
- Verify MiniMax H3 license terms for your region
FAQ
Q: What GPU do I need to run MiniMax H3 in ComfyUI?
An NVIDIA RTX 3060 with sufficient system RAM is the minimum requirement. For optimal performance, an RTX 3090 or 4090 is recommended, especially when working with multiple reference inputs at higher resolutions. The pruned int8 model variant works well on lower-VRAM cards.
Q: Can MiniMax H3 generate audio along with video?
Yes, MiniMax H3 generates native stereo audio including dialogue, sound effects, and music together with the video in a single MP4 file. You can also provide reference audio clips in R2V mode to clone voices or match specific audio styles.
Q: How many reference inputs can I use in Reference-to-Video mode?
R2V supports up to 9 reference images, 3 reference videos (each with its own soundtrack), and 3 standalone reference audio clips. You can mix and match these inputs to lock character identity, style, motion, camera moves, and voices.
Q: Is it possible to exceed the 15-second maximum video duration?
While the official maximum is 15 seconds, testing shows that setting the length to 20 seconds can still produce coherent results. The duration input snaps to the model's 17-frame-per-block grid at 24fps, so durations follow a specific increment pattern.
Q: How do I switch between Text-to-Video and Reference-to-Video in one workflow?
Use an RG3 Fast Group Bypass node to ensure only one model loader group is active at a time. A boolean switch based on the label set in the model loader groups lets you toggle between FL2VA and Ref2VA conditioning and latents easily.