- MiniMax H3 in ComfyUI: Run the AI video model locally with modular, customizable node workflows
- Hardware baseline: NVIDIA RTX 3060 and up with sufficient system RAM for reliable generation
- Multiple input modes: Supports text-to-video, first/last frame, and multi-reference image/video inputs
- Audio generation: Model can produce synchronized audio, voice cloning, and background music
- License requirement: Users in the USA and EU must request access via a license form before downloading
MiniMax H3 ComfyUI Overview
Running MiniMax H3 in ComfyUI gives you full local control over AI video generation. Instead of relying on cloud APIs, you can build modular node-based workflows that process text prompts, images, and reference videos entirely on your own hardware. The model handles complex prompts, multi-image references, audio synthesis, and even voice cloning.
Video Highlights:
- Text-to-video generation at 864x480 resolution using pruned int8 models
- First and last frame image inputs for controlled video transitions
- Multi-reference inputs combining up to four images, video clips, and audio tracks
- Voice cloning by merging audio from two separate video sources
- Workflow architecture using RG3 fast group bypass nodes for easy model switching
The core advantage of using ComfyUI for MiniMax H3 is flexibility. You can toggle between the first-last frame model and the reference model using a simple boolean switch. The K sampler and video output nodes follow standard ComfyUI conventions, making the workflow easy to modify and expand.
For NVIDIA RTX 3090 users, the pruned int8 model (FL2va) offers an excellent balance of quality and performance. If you have a 3060 or 3070, ComfyUI's optimization should still handle the workload as long as you have sufficient system RAM.
Hardware Requirements and Model Variants
MiniMax H3 is designed to run on consumer-grade NVIDIA GPUs. ComfyUI has confirmed that the RTX 3060 and above are supported, though generation times vary significantly based on your hardware tier.
| GPU Model | Minimum RAM | Expected Performance | Recommended Model |
|---|---|---|---|
| RTX 3060 | 32GB system | Slower generation, basic resolutions | Pruned int8 |
| RTX 3070/3080 | 32GB system | Moderate speed, standard resolutions | Pruned int8 |
| RTX 3090/4090 | 64GB system | Fast generation, higher resolutions | Pruned int8 or full |
Generation time depends heavily on input complexity. A simple text-to-video prompt at 864x480 may complete quickly, while multi-reference workflows with video and audio inputs can take five minutes or longer on an RTX 3090.
The model officially supports up to 15 seconds of video. However, testing shows that pushing to 20 seconds can still produce coherent output without garbled results. Treat this as experimental rather than guaranteed.
Input Modes and Generation Capabilities
MiniMax H3 offers several distinct input modes, each suited to different creative goals. Understanding when to use each mode helps you get the best results.
Text-to-Video
- No input images required
- Handles complex, detailed prompts
- Best for pure creative generation
- Supports illustrated and realistic styles
First & Last Frame
- Defines start and end frames
- Prompt controls transition style
- Use picture zero and picture one tags
- Video always starts and ends on provided images
Reference Model
- Accepts multiple images and videos
- Images guide appearance, not exact frames
- Supports up to 6+ combined inputs
- Enables voice cloning and audio merging
Resolution Options
The workflow includes a resolution selector with reference previews for all available output sizes.
| Resolution | Aspect Ratio | Use Case | Notes |
|---|---|---|---|
| 864x480 | 16:9 | Standard video | Fastest generation time |
| 1216x672 | 16:9 | Higher quality | Longer processing, sharper detail |
| Custom | Varies | Special formats | Adjust width/height in conditioning node |
MiniMax H3 generates synchronized audio alongside video. It can produce background music, character dialogue, and even clone voices from reference video clips. When using two input videos with separate audio tracks, the model can merge both voices into a newly generated scene.
Step-by-Step Workflow Setup
Building a MiniMax H3 workflow in ComfyUI requires careful organization of model loaders, conditioning nodes, and input selectors. Follow these steps to create a functional workflow.
Load the Model
Place the MiniMax H3 model file (such as pruned int8 FL2va) into your ComfyUI models directory. Use the dedicated model loader node for either the first-last frame version or the reference video version. Both require the same CLIP and VAE models.
Configure the RG3 Bypass
Add an RG3 fast group bypass node to ensure only one model group is active at a time. This node acts as a toggle, preventing conflicts between the first-last frame conditioning and the reference video conditioning.
Set Up Conditioning
For first-last frame mode, connect your first and last frame images to the conditioning node along with prompt text, width, height, and video length parameters. For reference mode, connect all image, video, and audio inputs to the reference-to-video node.
Configure the K Sampler
Use a standard K sampler node. No special configuration is needed beyond the usual steps, CFG, and denoise settings. The sampler processes the conditioned latents from either model path.
Set Output Options
Configure the video output node with your preferred prefix and format. Options include MP4 and animated GIF. Save the final image and audio separately if you plan to do frame interpolation later using FILM or RIFE models.
Use a boolean switch linked to the label set in the model loader groups. This ensures the correct conditioning and latents are sent through the pipeline when you toggle between the first-last frame and reference models.
Prompting Techniques and Reference Indexing
Effective prompting in MiniMax H3 workflows requires understanding how the model interprets image and video references. The reference model uses triangular bracket indexing to identify different input sources.
Prompt Reference Syntax
| Syntax | Purpose | Example Usage |
|---|---|---|
| [picture zero] | References first image input | Use for character appearance guide |
| [picture one] | References second image input | Use for secondary character or object |
| [video zero] | References first video input | Use for motion or scene reference |
| [video one] | References second video input | Use for transition target or style |
| [audio zero] | References audio-only input | Use for voice or music guide |
The indexing system is flexible. You can start from zero or one depending on your preference. The model does not enforce strict numbering, but consistency within a single workflow is recommended.
When using first and last frame images, write a transformational prompt that describes what happens between the two images. Mention picture zero near the beginning and picture one toward the end. While the model will always start and end on the provided frames regardless of prompt content, the prompt dictates when and how the transition occurs.
Advanced Multi-Input Example
The reference model can handle complex combinations of inputs. One tested configuration used four input images of different styles and sizes, one video clip, and the audio from that video, totaling six different inputs. The prompt instructed the model to create four characters in a room and end with the final video clip appearing as if viewed through a mirror.
In reference mode, images serve as appearance guides rather than exact start or end frames. The model uses them to determine how characters and objects should look, but the generated video will not necessarily begin or end with those images. This differs from first-last frame mode where the provided images are guaranteed as the start and end.
Licensing and Post-Processing
License Considerations
The MiniMax H3 license currently restricts downloads in the USA and European Union. However, individuals can request access by submitting a license request form. Based on user reports as of 2026, the approval email arrives quickly after submission.
This guide does not constitute legal advice. The license terms may change over time. Always read the official license agreement and conduct your own research before downloading and using the model. Commercial use may carry additional restrictions.
Frame Interpolation
After generating your video, you can improve smoothness using frame interpolation models available in ComfyUI.
| Interpolation Model | Output FPS | Multiplier | Quality |
|---|---|---|---|
| FILM | 48 fps | 2x | Smooth, natural motion |
| RIFE | 48 fps | 2x | Sharp, detailed frames |
| RIFE | 72 fps | 3x | Very smooth, longer processing |
Pre-Generation Checklist:
- Verify GPU has sufficient VRAM for chosen resolution
- Confirm CLIP and VAE models are loaded for both model paths
- Test RG3 bypass toggle switches between models correctly
- Save final image and audio separately for post-processing
- Check video output format matches your intended use
FAQ
Q: What GPU do I need to run MiniMax H3 in ComfyUI?
ComfyUI has confirmed support for NVIDIA RTX 3060 and above. You need sufficient system RAM alongside your GPU. An RTX 3090 with 64GB system RAM provides fast generation times, while a 3060 with 32GB RAM can still produce results at standard resolutions.
Q: Can MiniMax H3 generate audio with video?
Yes. The model generates synchronized audio including background music, dialogue, and sound effects. It also supports voice cloning by using audio tracks from reference video clips. You can merge voices from two separate videos into a single new scene.
Q: What is the maximum video length for MiniMax H3?
The official maximum is 15 seconds. However, testing shows that pushing to 20 seconds can still produce coherent output. Going beyond 15 seconds should be treated as experimental, as results may vary depending on prompt complexity and input types.
Q: How do I switch between first-last frame and reference models in one workflow?
Use an RG3 fast group bypass node to ensure only one model group is active at a time. Add a boolean switch linked to the label set in the model loader groups. This routes the correct conditioning and latents through the pipeline based on which model you select.
Q: Can I use more than two reference images?
Yes. The reference model supports multiple images, videos, and audio-only inputs. Tested configurations include four images plus one video with audio, totaling six inputs. Expanding to five or six images is straightforward by adding additional input nodes to the workflow.