MiniMax H3 ComfyUI

INITIAL RELEASE

This is the initial release of MiniMax H3 ComfyUI integration, providing
full local deployment of the MiniMax H3 omni-modal generative model within
the ComfyUI node-based interface.

INCLUDED NODES

• H3ModelLoader: Load MiniMax H3 model (FL2VA or Ref2VA variant)
• H3TextToVideo: Text prompt to video+audio generation
• H3ImageToVideo: First/last frame image input to video generation
• H3ReferenceToVideo: Multi-modal reference (images, videos, audio) to video
• H3VAEDecode: Decode H3 visual and audio latents to video and stereo audio
• H3ContextIR: Call H3-Context-IR API for prompt enhancement
• H3Regenerate2K: Call H3-Regenerate-2K API for 2K resolution upscaling

INCLUDED WORKFLOWS

• t2v_workflow.json: Text-to-video basic workflow
• i2v_workflow.json: Image-to-video (first frame) workflow
• r2v_workflow.json: Reference-to-video workflow
• t2v_2k_workflow.json: Text-to-video with 2K regeneration
• audio_video_workflow.json: Combined audio+video decoding workflow
• batch_generation.json: Batch multi-prompt generation

INCLUDED SCRIPTS

• download_model.py: Download MiniMax H3 from HuggingFace
• sglang_serve.py: Deploy H3 with SGLang inference server
• inference_t2v.py: Standalone text-to-video inference
• inference_i2v.py: Standalone image-to-video inference
• inference_r2v.py: Standalone reference-to-video inference
• api_2k_workflow.py: Full 2K workflow (local H3 + API modules)
• convert_to_comfyui.py: Convert HF checkpoint to ComfyUI format

MODEL DETAILS

• Architecture: H3-Omni-Transformer (33B dense, single-stream)
• Text Encoder: H3-Encoder (Qwen3-VL-32B based, layer 50 hidden states)
• Visual VAE: H3-VisualVAE (f16t4d24, 24 latent channels)
• Audio VAE: H3-AudioVAE (32 kHz stereo, 40 Hz latent rate)
• Position Encoding: MM-RoPE (3D Multimodal Rotary Position Embeddings)
• Precision: BF16
• Inference Frameworks: SGLang, vLLM, diffusers, ComfyUI

OUTPUT SPECIFICATIONS

• Duration: 4-15 seconds
• Resolution: 768p native (2K via H3-Regenerate-2K API)
• Frame Rate: 24 FPS
• Audio: 32 kHz stereo
• Aspect Ratios: 21:9, 16:9, 4:3, 1:1, 3:4, 9:16
• Languages: 11 stable (Arabic, Chinese, English, French, German, Italian,
  Japanese, Korean, Portuguese, Russian, Spanish)

COMPATIBILITY

• ComfyUI v0.31.0 (released August 8, 2026)
• ComfyUI v0.30.0 (H3 support added in this version)
• PyTorch 2.1+ with CUDA 12.0+
• Python 3.10+
• 4x NVIDIA GPU (80GB VRAM each recommended)

Full Changelog: https://github.com/AyushS20/MiniMax-H3-ComfyUI/commits/MiniMaxH3