MiniMax H3 - Image to Video
ComfyUI in the Cloud
Deploy MiniMax H3 - Image to Video, from
$0.60 / hour
MiniMax-H3 FL2VA Video Generation ComfyUI Workflow
A ComfyUI video generation workflow powered by MiniMax-H3, featuring an INT8-quantized diffusion model, Qwen3-VL 32B vision-language conditioning, dedicated video and audio autoencoders, and optional Turbo LoRA acceleration.
Workflow Overview
This workflow utilizes the MiniMax-H3 FL2VA model configuration for AI video generation within ComfyUI. It combines an INT8-quantized diffusion model, Qwen3-VL 32B vision-language conditioning, a dedicated video VAE, and a separate audio VAE.
The workflow exposes configurable video width, height, duration, and noise seed parameters. Width and height are connected as external inputs, allowing other nodes to determine the video dimensions. The duration parameter is configured to 5.0.
An optional Turbo LoRA configuration is included for accelerated generation. The selected Turbo model is an 8-step BF16 variant, configured with a model strength of 1.00. Turbo mode is currently disabled, so these optional acceleration settings are not active in the displayed configuration.
The diffusion model and video VAE use INT8 quantization to help reduce memory requirements. The audio VAE uses FP32 precision, while the Qwen3-VL 32B conditioning model includes NVFP4 and AWQ quantization identifiers.
Generation Settings
Width
External Input
Height
External Input
Duration
5.0
Noise Seed
757358688076805
Turbo Mode
Disabled
Turbo Steps
8
Turbo Strength
1.00
Model Precision
INT8
Model Configuration
Video Generation Model (UNet)
INT8minimax_h3_fl2va_pruned_int8_convrot
The MiniMax-H3 diffusion model serves as the primary video generation component. It performs the denoising process responsible for producing video latent representations. The selected model uses a pruned INT8-quantized configuration to help reduce memory requirements compared with higher-precision model weights. The model can receive conditioning information from compatible text and visual processing nodes.
Text / Vision Conditioning Model
Qwen3-VL 32Bqwen3vl_32b_minimax_h3_nvfp4_awq
The Qwen3-VL 32B model provides vision-language conditioning for the MiniMax-H3 generation process. It supports processing text instructions and visual information into representations that guide the diffusion model. The selected model identifier includes NVFP4 and AWQ quantization designations. This component can be integrated with compatible prompt-processing and visual conditioning nodes within the larger workflow.
Video Variational Autoencoder (VAE)
INT8minimax_h3_video_vae_int8_convrot
The Video VAE provides the autoencoder component used for video latent processing. During decoding, the VAE converts generated latent representations into visible video frames. The selected INT8-quantized configuration helps reduce memory requirements. The decoded frames can be passed to compatible video processing, preview, or export nodes.
Audio Variational Autoencoder (VAE)
FP32minimax_h3_audio_vae_fp32
The Audio VAE provides a dedicated component for handling audio latent representations. Audio autoencoders support encoding and decoding between audio information and latent representations. The selected model uses FP32 precision. This component can be connected to compatible audio processing and decoding nodes within the complete video generation workflow.
Optional Turbo LoRA
Disabledminimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16
The workflow includes an optional MiniMax-H3 Turbo LoRA for accelerated video generation. The selected model is identified as an 8-step Turbo variant using BF16 precision. The configured LoRA model strength is 1.00. Turbo mode is disabled in the displayed configuration, meaning the optional acceleration mode is not currently enabled. When activated through the appropriate workflow connections, the Turbo LoRA can modify the diffusion model's behavior.
Turbo LoRA Configuration
The workflow includes a configurable Turbo mode intended for accelerated video generation. The selected Turbo LoRA is an 8-step BF16 model with a configured strength of 1.00. The Turbo mode toggle is currently disabled.
| Turbo Mode | Disabled |
| Turbo LoRA | minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16 |
| Turbo Model Strength | 1.00 |
| Turbo Steps | 8 |
| LoRA Precision | BF16 |
| LoRA Version | v1.0 |
The Turbo strength parameter controls the configured influence of the selected LoRA when applied. A strength value of 1.00 represents the full configured LoRA weight. The Turbo step value is set to 8, but because Turbo mode is disabled, this does not establish the actual sampling step count used by the workflow.
Video Generation Configuration
The workflow provides width, height, and duration parameters for configuring video generation. Width and height are exposed as external input connections, allowing upstream nodes to control the video dimensions. The displayed duration value is 5.0.
Video Width
External Input
Determines the horizontal dimension of the generated video through a connected node.
Video Height
External Input
Determines the vertical dimension of the generated video through a connected node.
Video Duration
5.0
Defines the configured video duration used by the generation workflow.
The duration value corresponds to a five-second configuration when interpreted in seconds by the internal workflow nodes. The actual frame rate and frame count are not visible in the screenshot. Video resolution depends on the connected width and height inputs.
Noise Seed Configuration
The workflow includes a configurable noise seed that controls random initialization during the video generation process. The displayed seed value is 757358688076805.
| Noise Seed | 757358688076805 |
| Seed Control | Configurable |
| Randomization | Seed adjustment supported |
Using a fixed seed can help reproduce generation results when model configuration, conditioning, and other workflow parameters remain unchanged. Changing the seed may produce different video variations by modifying the random initialization used during generation.
Video and Audio Processing
The workflow includes separate video and audio autoencoders for processing their respective latent representations. These components allow compatible downstream nodes to perform video and audio decoding.
Video Autoencoder
minimax_h3_video_vae_int8_convrot
Precision: INT8
Handles video latent representations and supports decoding generated video data into visible frames. INT8 quantization helps reduce memory usage.
Audio Autoencoder
minimax_h3_audio_vae_fp32
Precision: FP32
Provides encoding and decoding capabilities for compatible audio processing components. The FP32 configuration uses 32-bit floating-point precision.
The inclusion of separate audio and video autoencoders provides the workflow with components for processing both modalities. The exact audio generation and synchronization behavior depends on the internal workflow nodes and their connections.
Surrounding Nodes and Workflow Integration
The MiniMax-H3 FL2VA subgraph consolidates diffusion model selection, conditioning models, audio and video autoencoders, generation parameters, and optional Turbo settings into a reusable component.
The workflow exposes width and height input connections for controlling video dimensions. These inputs can receive values from upstream nodes responsible for resolution configuration. The duration and noise seed parameters control the configured generation length and random initialization.
In a typical MiniMax-H3 workflow, text prompts and visual inputs are processed through compatible conditioning nodes. The Qwen3-VL 32B model provides representations that can guide the MiniMax-H3 diffusion model.
The diffusion model performs the denoising process using the configured generation parameters. The optional Turbo LoRA can modify this process when Turbo mode is enabled. In the displayed configuration, Turbo mode is disabled.
After the generation process, the video VAE can decode video latent representations into visible frames. The audio VAE provides a separate component for compatible audio decoding operations.
The generated video and audio data can be connected to downstream processing or export nodes. Exact prompt connections, reference inputs, sampler settings, frame rate, and output nodes are not visible in the supplied screenshot.
Workflow Highlights
- ✓ MiniMax-H3 FL2VA: Uses the MiniMax-H3 diffusion model for AI video generation.
- ✓ INT8 Quantization: Uses quantized diffusion and video VAE components to help reduce memory requirements.
- ✓ Qwen3-VL 32B: Integrates vision-language conditioning with NVFP4 and AWQ quantization identifiers.
- ✓ Configurable Duration: Includes a video duration parameter set to 5.0.
- ✓ Optional Turbo Mode: Supports an 8-step Turbo LoRA configuration, currently disabled.
- ✓ Turbo LoRA Strength: Includes an adjustable model strength configured to 1.00.
- ✓ Video Autoencoder: Uses an INT8 video VAE for decoding generated video latent representations.
- ✓ Audio Autoencoder: Includes an FP32 audio VAE for compatible audio processing operations.
- ✓ Flexible Video Dimensions: Width and height are supplied through external node connections.
- ✓ Seed-Based Generation: Includes configurable noise initialization with the displayed seed set to 757358688076805.
- ✓ Reusable Subgraph: Organizes model configuration, generation parameters, and optional acceleration settings for integration into larger ComfyUI workflows.
Workflow Summary
This MiniMax-H3 FL2VA ComfyUI workflow combines an INT8-quantized diffusion model with Qwen3-VL 32B conditioning, an INT8 video VAE, and an FP32 audio VAE. It includes externally connected width and height inputs, a duration setting of 5.0, and a noise seed of 757358688076805.
The workflow also includes an optional MiniMax-H3 Turbo LoRA configured as an 8-step BF16 variant. The LoRA strength is set to 1.00, and the Turbo steps parameter is set to 8. Turbo mode is currently disabled.
Qwen3-VL 32B provides vision-language conditioning, while the MiniMax-H3 diffusion model performs the video generation process. Dedicated video and audio autoencoders support processing and decoding their respective latent representations.
The subgraph provides a reusable configuration for MiniMax-H3 video generation, combining model selection, duration, noise initialization, optional Turbo acceleration, and configurable video dimensions within ComfyUI.