MiniMax H3 - Reference to Video
ComfyUI in the Cloud
Deploy MiniMax H3 - Reference to Video, from
$0.60 / hour
MiniMax-H3 Reference-to-Video ComfyUI Workflow
A ComfyUI video generation workflow configuration powered by MiniMax-H3, featuring an INT8-quantized diffusion model, a 4-step Turbo LoRA, Qwen3-VL 32B conditioning, and dedicated video and audio autoencoders.
Workflow Overview
This workflow uses the MiniMax-H3 model configuration for AI video generation within ComfyUI. It combines an INT8-quantized diffusion model, a Turbo LoRA, Qwen3-VL 32B conditioning, and separate video and audio autoencoders.
The selected diffusion model is identified as a reference-to-video-oriented model. The additional Turbo LoRA is designed for a 4-step generation configuration, providing an optional acceleration component for the underlying diffusion model.
The workflow also includes dedicated video and audio VAE models for processing their respective latent representations. INT8 quantization is used for the primary diffusion model and video VAE, while the audio VAE uses FP32 precision. The displayed nodes form the model-loading section of a larger ComfyUI workflow.
Model Configuration
Diffusion Model
MiniMax-H3
Diffusion Precision
INT8
LoRA Strength
1.00
Turbo LoRA Variant
4-Step
Conditioning Model
Qwen3-VL 32B
CLIP Type
minimax
Video VAE
INT8
Audio VAE
FP32
Included Nodes
Load Diffusion Model
INT8minimax_h3_ref2va_pruned_int8_convrot
The Load Diffusion Model node loads the primary MiniMax-H3 diffusion model used for video generation. The selected model is a pruned, INT8-quantized configuration identified for reference-to-video generation. INT8 quantization helps reduce model memory requirements compared with higher-precision weights. The node exposes a MODEL output that can be connected to additional model-processing nodes, including the Turbo LoRA loader.
Load LoRA
Turbo / BF16minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16
The Load LoRA node applies an additional MiniMax-H3 Turbo LoRA configuration to the loaded diffusion model. The selected LoRA is identified as a reference-to-video Turbo variant designed for 4-step generation. The LoRA model uses a BF16 configuration and is applied with a model strength of 1.00. The node receives the diffusion model through its MODEL input and exposes the modified model through its MODEL output.
Load CLIP / Vision-Language Model
Qwen3-VL 32Bqwen3vl_32b_minimax_h3_nvfp4_awq
The Load CLIP node loads the Qwen3-VL 32B conditioning model used with MiniMax-H3. This vision-language model is intended to process text and visual information and provide conditioning representations for the generation process. The selected model identifier includes NVFP4 and AWQ, indicating a quantized model configuration. The node uses the minimax model type and exposes a CLIP output for connection to compatible conditioning nodes.
Load Video VAE
INT8minimax_h3_video_vae_int8_convrot
The Video VAE provides the autoencoder component used for processing video latent representations. In a typical video generation workflow, the video decoder converts generated latent representations into visible video frames. This selected VAE uses an INT8-quantized configuration to help reduce memory usage. The node exposes a VAE output that can be connected to compatible video encoding or decoding nodes.
Load Audio VAE
FP32minimax_h3_audio_vae_fp32
The Audio VAE provides a separate autoencoder component for processing audio latent representations. Audio autoencoders convert audio information between waveform-oriented representations and compressed latent representations. The selected model uses FP32 precision, providing a higher-precision configuration than the INT8 video VAE. The node exposes a VAE output for use with compatible audio processing and decoding components.
Turbo LoRA Configuration
The workflow includes a MiniMax-H3 Turbo LoRA intended for accelerated reference-to-video generation. The model identifier specifies a 4-step Turbo variant with version 0.1. The displayed configuration applies the LoRA with a model strength of 1.00.
| LoRA Model | minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16 |
| LoRA Strength | 1.00 |
| Turbo Variant | 4-Step |
| LoRA Precision | BF16 |
| LoRA Version | v0.1 |
The LoRA loader receives the selected MiniMax-H3 diffusion model and applies the configured model modification. A strength value of 1.00 represents the full configured LoRA weight. The 4-step designation comes from the LoRA model name. The actual sampling step count is not visible in the provided screenshot.
Vision-Language Conditioning
The workflow uses a Qwen3-VL 32B conditioning model configured for the MiniMax-H3 architecture. The selected CLIP loader uses the minimax type, allowing the loaded model to be used by compatible conditioning nodes.
| Conditioning Model | Qwen3-VL 32B |
| Model Identifier | qwen3vl_32b_minimax_h3_nvfp4_awq |
| CLIP Type | minimax |
| Quantization | NVFP4 / AWQ |
Vision-language conditioning can provide representations derived from text instructions and visual reference information. In a reference-to-video workflow, compatible conditioning nodes can use these representations to guide the video generation process. The exact conditioning inputs and processing nodes are not visible in the screenshot.
Video and Audio Autoencoder Configuration
The workflow includes separate VAE models for video and audio processing. These components provide dedicated latent representation handling for their respective data types.
Video Autoencoder
minimax_h3_video_vae_int8_convrot
Precision: INT8
Handles video latent representations and supports decoding generated video data into visible frames. The INT8 configuration helps reduce memory requirements.
Audio Autoencoder
minimax_h3_audio_vae_fp32
Precision: FP32
Handles audio latent representations and provides encoding or decoding capabilities for compatible audio processing nodes. The FP32 configuration uses 32-bit floating-point precision.
Separate audio and video autoencoders allow the workflow to support components for both modalities. The exact audio generation, video decoding, and synchronization behavior depends on the connected nodes in the complete workflow.
Additional Parameters
| Diffusion Model | MiniMax-H3 Ref2VA |
| Diffusion Precision | INT8 |
| Turbo LoRA | Ref2V 4-Step v0.1 |
| LoRA Strength | 1.00 |
| CLIP Type | minimax |
| Conditioning Model | Qwen3-VL 32B |
| Video VAE Precision | INT8 |
| Audio VAE Precision | FP32 |
| Sampling Settings | Not visible in the screenshot |
| Video Resolution | Not visible in the screenshot |
Surrounding Nodes and Workflow Integration
The displayed MiniMax-H3 workflow section contains five model-loading nodes. These nodes prepare the diffusion model, Turbo LoRA, vision-language conditioning model, video VAE, and audio VAE for use by other components in the ComfyUI graph.
The Load Diffusion Model node provides the primary MiniMax-H3 model through its MODEL output. This output is connected to the Load LoRA node, where the selected 4-step Turbo LoRA is applied with a strength value of 1.00. The modified MODEL output can then be passed to compatible video generation or sampling nodes.
The Load CLIP node independently loads the Qwen3-VL 32B conditioning model. Its CLIP output can be connected to compatible text and visual conditioning nodes. In a typical reference-to-video workflow, those nodes prepare conditioning information derived from prompts and reference images.
The two Load VAE nodes provide separate autoencoder components for video and audio. The Video VAE can be used by downstream video encoding or decoding nodes, while the Audio VAE can be connected to compatible audio processing components.
Together, these model-loading nodes establish the primary components needed by a larger MiniMax-H3 video generation workflow. Reference image inputs, prompt processing, sampling configuration, video decoding, audio processing, and final output handling depend on the surrounding ComfyUI nodes.
Workflow Highlights
- ✓ MiniMax-H3: Uses the MiniMax-H3 diffusion model configuration for reference-to-video-oriented generation.
- ✓ INT8 Quantization: Uses an INT8 diffusion model to help reduce memory requirements.
- ✓ Turbo LoRA: Includes a 4-step Turbo LoRA variant configured at strength 1.00.
- ✓ Qwen3-VL 32B: Integrates a vision-language conditioning model with NVFP4/AWQ quantization identifiers.
- ✓ Video Autoencoder: Includes an INT8 video VAE for processing video latent representations.
- ✓ Audio Autoencoder: Includes a separate FP32 audio VAE for compatible audio processing operations.
- ✓ Modular Model Loading: Uses separate loader nodes for the diffusion model, LoRA, conditioning model, and autoencoders.
- ✓ Workflow Integration: Exposes MODEL, CLIP, and VAE outputs for integration into larger ComfyUI graphs.
Workflow Summary
This ComfyUI workflow section combines the MiniMax-H3 INT8 diffusion model with a reference-to-video Turbo LoRA, Qwen3-VL 32B conditioning, and dedicated video and audio autoencoders.
The selected Turbo LoRA is identified as a 4-step variant and is configured with a model strength of 1.00. The diffusion model and video VAE use INT8 quantization, while the audio VAE uses FP32 precision.
The Qwen3-VL 32B model provides the vision-language conditioning component, while the separate video and audio VAEs support processing their respective latent representations. The displayed nodes form a modular model-loading configuration intended for integration into a larger MiniMax-H3 video generation workflow.