ComfyUI-Open-H3-Regenerate-2K

ComfyUI-Open-H3-Regenerate-2K

A dedicated ComfyUI custom node engineered for NVIDIA's Open-H3-Regenerate-2K video super-resolution pipeline on MiniMax H3. It integrates 2K latent refinement into a single unified node with automatic weight downloading, SwiGLU-fixed native BF16 support, 50 LMS Sharp AdaLN curve branch injection, calibrated Euler 2/3-step sigmas, and audio-locked latent masking.

<img width="1373" height="686" alt="image" src="https://github.com/user-attachments/assets/8f794f1a-6c45-429e-bcc4-86741276fec5" />

News & Updates

• 2026/10/28: Release ComfyUI-Open-H3-Regenerate-2K v1.0.0 ( update.md )
  • Initial release of Open-H3 Regenerate 2K (AILab_OpenH3_Regenerate_2K) node.
  • Built-in automatic model download system fetching weights directly to ComfyUI/models/loras/minmax/.
  • Added SwiGLU-fixed native BF16 (8.80 GB) and lightweight Rank-128 BF16 (1.95 GB) models.
  • Built-in 50 LMS Sharp AdaLN curve branches (sharp_curve) and audio-locking (lock_audio).

Features

• All-in-One 2K Refinement: Directly outputs patched MODEL, audio-locked LATENT, and official 2/3-step SIGMAS in a single pass.
• Automatic Model Download: Missing weights automatically download to ComfyUI/models/loras/minmax/ on first run.
• Flexible Directory Lookup: Recursively searches models/loras/ so users can organize models in any subfolder without re-downloading.
• SwiGLU-Fixed BF16 Native Weights: Pre-fused QKV and aligned SwiGLU channels for ComfyUI H3 base architecture.
• 50 LMS Sharp AdaLN Curve Branches: Built-in support for Sharp curve conditioning to enhance 2K video textures.
• Audio Stream Protection: Built-in lock_audio masks audio latents at sigma 0 so soundtrack fidelity is fully preserved.

Installation

Method 1: Install via ComfyUI Manager (Recommended)

Search for ComfyUI-Open-H3-Regenerate-2K in ComfyUI Manager and click Install.

Method 2: Manual Clone

cd ComfyUI/custom_nodes
git clone https://github.com/1038lab/ComfyUI-Open-H3-Regenerate-2K.git
pip install -r ComfyUI-Open-H3-Regenerate-2K/requirements.txt

Model Downloads

[!TIP] Automatic Download Enabled: You do not need to download Open-H3 LoRA weights manually. The node downloads them on-demand to ComfyUI/models/loras/minmax/. Direct download links are provided below for offline environments.

1. Open-H3 2K LoRA Weights (Target: ComfyUI/models/loras/minmax/)

• Open-H3-Regenerate-2K Full-Rank BF16 (8.80 GB, Recommended): Download open_h3_regenerate_2k_bf16.safetensors
• Open-H3-Regenerate-2K r128 BF16 (1.95 GB, Fast / Low VRAM): Download minimax_h3_open_regenerate_2k_r128_bf16.safetensors
• Sharp Curve Branches (1.24 GB, 50 AdaLN branches): Download sharp_curve_branches.safetensors

2. MiniMax H3 Base Models (Required for Workflow)

Place these standard MiniMax H3 models in their respective ComfyUI directories:

• Diffusion Model (ComfyUI/models/diffusion_models/): minimax_h3_ref2va_pruned_int8_convrot.safetensors
• Text Encoder (ComfyUI/models/text_encoders/): qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
• Video VAE (ComfyUI/models/vae/): minimax_h3_video_vae_fp16.safetensors
• Audio VAE (ComfyUI/models/vae/): minimax_h3_audio_vae_fp32.safetensors
• 3D Latent Upscaler (ComfyUI/models/latent_upscale_models/): minimax_h3_latent_upscaler_3d_conv_v1_bf16.safetensors

Usage

Node: Open-H3 Regenerate 2K (AILab_OpenH3_Regenerate_2K)

Category: 🧪AILab/Open-H3

1. Connect the MiniMax H3 base model to model.
2. Connect the 3D-upscaled joint audio-video latent to latent.
3. Choose your preset in model_name:
   • Open-H3-Regenerate-2K (BF16) for highest 2K quality.
   • Open-H3-Regenerate-2K-r128 (BF16) for faster generation and lower memory.
4. Outputs:
   • MODEL: Patched model ready for sampling.
   • LATENT: Latent with audio stream protected at sigma 0.
   • SIGMAS: Official calibrated Euler sigmas (3-step or 2-step) to connect to SamplerCustomAdvanced.

Recommended Settings

• Sampler: euler
• CFG: 1.0
• Steps: "3" (official schedule: 0.909375 -> 0.857143 -> 0.666667 -> 0.0)
• Strength: 1.0
• Sharp Curve: True
• Lock Audio: True

Workflows

License & Acknowledgements

• Model Hosting & Conversion: 1038lab/Open-H3-Regenerate-2K
• Upstream Research: NVIDIA / Efficient-Large-Model — Open-H3-Regenerate-2K, MiniMaxAI/MiniMax-H3, Comfy-Org/MiniMax-H3, LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler.
• License: Code released under GPL-3.0 License.