RunningHub MiniMax H3

ComfyUI-RH-MiniMax-H3

Native MiniMax-H3 audio-video generation nodes for ComfyUI. Model components run inside the ComfyUI process without an SGLang server or Diffusers pipeline. Based on the official MiniMax-H3 project.

✨ Features

• T2VA, FL2VA, Ref2VA, and video-to-audio generation
• A focused UI with two generation nodes and three model loaders
• INT8 ConvRot weights, optional turbo LoRA, offload, and attention backends
• Attention backends: auto, sdpa, sage, and ck (Comfy Kitchen INT8, same kernel as --use-ck-attention)
• Example workflows for text, keyframe, multimodal-reference, and V2A tasks

🛠️ Installation

cd ComfyUI/custom_nodes
git clone https://github.com/RH-RunningHub/ComfyUI-RH-MiniMax-H3.git
pip install -r ComfyUI-RH-MiniMax-H3/requirements.txt

Restart ComfyUI after installation.

📦 Model Download & Installation

Use the complete converted INT8 ConvRot model package.

| Priority | Source | Purpose | |---|---|---| | 1 | Hugging Face INT8 ConvRot | Preferred converted weights | | 2 | ModelScope INT8 ConvRot | Preferred China mirror |

The converted repositories contain the same complete bundle. Run one of the following commands from the ComfyUI root directory. The download is about 95 GiB, so keep at least 110 GiB of free disk space.

Model Directory Structure

All models should be placed in ComfyUI/models/MiniMax-H3-INT8-CONVROT/ with the following structure:

ComfyUI/
└── models/
    └── MiniMax-H3-INT8-CONVROT/
        ├── MiniMax-H3-FL2VA-int8_convrot.safetensors
        ├── MiniMax-H3-Ref2VA-int8_convrot.safetensors
        ├── qwen3-vl-32b-int8_convrot.safetensors
        ├── MiniMax-H3-video_vae.safetensors
        ├── MiniMax-H3-audio_vae.safetensors
        ├── minimax_h3_fl2v_turbo_4step_v0.1.safetensors
        ├── minimax_h3_fl2v_turbo_4step_v1.0_768p_bf16.safetensors
        ├── minimax_h3_fl2v_turbo_8step_v1.0_bf16.safetensors
        ├── minimax_h3_ref2v_turbo_4step_v0.1_bf16.safetensors
        ├── FL2VA/
        │   ├── model_index.json
        │   ├── transformer/config.json
        │   ├── text_encoder/config.json
        │   ├── tokenizer/
        │   ├── processor/
        │   ├── video_vae/config.json
        │   ├── video_vae/source/config.json
        │   └── audio_vae/config.json
        └── Ref2VA/
            └── ... same configuration layout

The plugin automatically detects the complete converted bundle. The legacy ComfyUI/models/MiniMax-H3 directory remains supported.

Download Methods

Method 1: Download from Hugging Face (Recommended)

cd /path/to/ComfyUI
python3 -m pip install -U huggingface_hub
hf download Gluttony10/MiniMax-H3-INT8-CONVROT \
  --local-dir ./models/MiniMax-H3-INT8-CONVROT

Re-run the same command to resume or update an interrupted download. On a high-bandwidth machine with at least 64 GiB RAM, prefix the hf download command with HF_XET_HIGH_PERFORMANCE=1 for maximum throughput.

Method 2: Download from ModelScope (For China users)

cd /path/to/ComfyUI
python3 -m pip install -U modelscope
modelscope download --model Gluttony10/MiniMax-H3-INT8-CONVROT \
  --local_dir ./models/MiniMax-H3-INT8-CONVROT

Method 3: Manual Download

| Model | Link | Description | |---|---|---| | Hugging Face bundle | Gluttony10/MiniMax-H3-INT8-CONVROT | Complete INT8 ConvRot package, VAEs, and turbo LoRAs | | ModelScope bundle | Gluttony10/MiniMax-H3-INT8-CONVROT | Same package for China users |

Optional Turbo LoRA

Select one turbo LoRA in the model loader when you want fewer sampling steps. Leave the LoRA empty to run the base converted weights.

| File | Typical use | |---|---| | minimax_h3_fl2v_turbo_4step_v0.1.safetensors | FL2VA / T2VA 4-step turbo | | minimax_h3_fl2v_turbo_4step_v1.0_768p_bf16.safetensors | FL2VA / T2VA 4-step turbo for 768p | | minimax_h3_fl2v_turbo_8step_v1.0_bf16.safetensors | FL2VA / T2VA 8-step turbo | | minimax_h3_ref2v_turbo_4step_v0.1_bf16.safetensors | Ref2VA 4-step turbo |

Model weights are not covered by this repository's Apache-2.0 license. Review the upstream model terms before use.

🚀 Usage

The main nodes are under RunningHub/MiniMax H3. Both generation nodes output frames, audio, and av_latent. Start with the bundled workflows:

• Text-to-video+audio
• Video-to-audio
• Image/audio reference generation

Older workflows can be migrated with:

python3 tools/migrate_workflow.py old_workflow.json --in-place

Legacy granular nodes remain registered so existing workflows still load and run, but they are deprecated and hidden from search, the node tree, and slot-drag suggestions.

📝 Node Reference

| Node | Purpose | |---|---| | RHMiniMaxH3ModelLoader | Load the FL2VA / Ref2VA DiT and optional LoRA | | RHMiniMaxH3TextEncoderLoader | Load the Qwen3-VL text encoder | | RHMiniMaxH3VAELoader | Load the video and audio VAEs | | RHMiniMaxH3VideoGen | T2VA / FL2VA / V2A from text, keyframes, or a source video | | RHMiniMaxH3RefGen | Ref2VA with ordered image, video, and audio references |

attention_backend on the loaders and generation nodes:

| Value | Behavior | |---|---| | auto | Follow ComfyUI's current optimized attention | | sdpa | PyTorch SDPA | | sage | SageAttention when available | | ck | Comfy Kitchen INT8 attention (same kernel as --use-ck-attention; errors if unavailable) |

Requirements

• ComfyUI 0.27 or newer (0.28+ recommended)
• A ComfyUI-compatible CUDA build of PyTorch, Triton, and comfy-kitchen
• ffmpeg and ffprobe for Ref2VA video/audio references
• Python packages listed in requirements.txt
• Sufficient RAM, VRAM, and fast model storage for MiniMax-H3

📄 License

Plugin code is licensed under Apache-2.0. Model weights use their upstream licenses. See NOTICE.md for attribution and third-party notices.

🔗 Links

• Official MiniMax-H3 project
• Hugging Face INT8 ConvRot weights
• ModelScope INT8 ConvRot weights

🙏 Acknowledgements

This project is based on MiniMax-H3, developed by MiniMax-AI.