slmonker/Open-H3-Regenerate-2K-IC-ComfyUI

Open-H3-Regenerate-2K-IC · ComfyUI LoRA

ComfyUI-format conversion of the official Open-H3-Regenerate-2K-IC IC LoRA for MiniMax-H3 transformer_ref.

The companion node pack and workflow are available at:

• GitHub: https://github.com/slmonker/ComfyUI-OpenH3Regenerate2KIC
• Upstream project: https://github.com/NVlabs/Sana/tree/sol-engine/models/Open-H3-Regenerate-2K/IC

File

minimax_h3_open_regenerate_2k_ic_comfy_bf16.safetensors

Place it in:

ComfyUI/models/loras/

Load order in ComfyUI:

1. minimax_h3_hyperflow_8step_v1.0_comfyui_bf16.safetensors
2. minimax_h3_lms_v1.0_r64.safetensors
3. minimax_h3_open_regenerate_2k_ic_comfy_bf16.safetensors (this file)

What was converted

The official file targets Diffusers/PEFT keys on MiniMax-H3 transformer_ref:

transformer.<module>.lora_A.weight
transformer.<module>.lora_B.weight

The converter maps it to ComfyUI's fused MiniMax-H3 layout:

• q/k/v → qkv_proj, with block-diagonal B
• ff.net.0.proj SwiGLU row order from Diffusers [value;gate] to ComfyUI [gate;value]
• fused qkv alpha = 3 × rank
• all other modules alpha = rank
• output dtype: BF16
• source scale remains 1.0

Conversion statistics:

Property	Value
Source linear layers	312
Fused ComfyUI modules	208
Tensors	624
Source rank	384
Base model	MiniMax-H3 transformer_ref
SHA256	a0c93dce50d9034cc084073b6c095d09ff25ad02351ac4d4f10113d67ae64ab3

Every target base-model weight exists in minimax_h3_ref2va_bf16.safetensors.

Text-to-video example

The companion GitHub repository now includes a complete T2VA workflow:

• English: https://github.com/slmonker/ComfyUI-OpenH3Regenerate2KIC/blob/main/README.md
• 中文: https://github.com/slmonker/ComfyUI-OpenH3Regenerate2KIC/blob/main/README.zh-CN.md
• Workflow: https://github.com/slmonker/ComfyUI-OpenH3Regenerate2KIC/blob/main/workflows/regen2k_ic_T2VA.json

It generates a 768P MiniMax-H3 draft from text, then refines the same clip to 2560×1440 with this IC LoRA.

Usage

Use the companion workflow:

open_h3_regenerate_2k_ic.json

It performs:

768P draft -> Ref2VA <Video 1> -> official 3D latent upscaler
           -> up3 sigma grid -> HyperFlow + LMS Sharp + this IC LoRA
           -> fast H3 VAE -> mux original draft audio

Default sampling grid:

up3: 0.834942, 0.696852, 0.468753, 0

License

This is a derivative adapter for MiniMax-H3. The model weights and this conversion are subject to the MiniMax-H3 Community License. The official LICENSE and upstream NOTICE files are included in this repository.

This is an independent community conversion and is not an official MiniMax product.