Extensions/Open-H3 Regenerate 2K IC
ComfyUI Extension

Open-H3 Regenerate 2K IC

ComfyUI nodes and workflows for Open-H3-Regenerate-2K-IC draft-to-2K video refinement

By slmonker·Created 2 days ago·Updated a day ago· 32
slmonker/ComfyUI-OpenH3Regenerate2KIC
Nodes—
On cloudLocal install
Stars32
Updateda day ago
Readme

ComfyUI Open-H3-Regenerate-2K-IC

English | 简体中文

ComfyUI nodes and example workflows for NVlabs/Sana · Open-H3-Regenerate-2K/IC.

This repository contains both:

  1. Text-to-video example — text → 768P H3 draft → IC 2K refinement.
  2. Draft-to-2K example — existing 768P draft → IC 2K refinement.

Demo videos

Each comparison shows the 768P draft on the left and the 2K IC result on the right. Click a preview to open the final MP4.

Demo 1

Demo 1

Draft MP4 · 2K MP4

Demo 2

Demo 2

Draft MP4 · 2K MP4

Demo 3

Demo 3

Draft MP4 · 2K MP4

Open the demo release

Examples

| Example | Workflow | Description | |---|---|---| | Text-to-video | workflows/regen2k_ic_T2VA.json | Generates a 1344×768 draft from text, then refines it to 2560×1440 | | Reference-to-video | workflows/H3_regen2k_ic_Ref2va.json | Enhances a reference image into a 768P Ref2VA draft, then refines it to 2560×1440 | | Draft-to-2K | workflows/open_h3_regenerate_2k_ic.json | Refines an existing 768P draft to 2560×1440 |

Text-to-video pipeline

Text prompt
  -> MiniMax-H3 Image-to-Video + DMAD 4-step LoRA
  -> 1344×768 / 124 frames / 24 fps draft
  -> decode draft video + draft audio
  -> draft becomes <Video 1> in Ref2VA
  -> official 3D latent upscaler
  -> HyperFlow + LMS Sharp + Open-H3 IC LoRA
  -> up3 3-step refinement
  -> 2560×1440 final video

The T2VA workflow uses one text node that feeds both stages:

  • raw text into the H3 draft generator
  • official enhancement prefix + the same text into the IC Ref2VA stage

Draft-to-2K pipeline

Existing 768P draft
  -> <Video 1> Ref2VA conditioning
  -> H3 VAE encode -> official 3D latent upscaler
  -> HyperFlow + LMS Sharp + IC LoRA
  -> up3 / up4 / hf8 sigma grid
  -> fast H3 VAE decode
  -> mux the draft's original audio

Contents

| Path | Purpose | |---|---| | nodes.py | Four ComfyUI helper nodes | | workflows/open_h3_regenerate_2k_ic.json | Draft-to-2K UI workflow | | workflows/open_h3_regenerate_2k_ic_api.json | Draft-to-2K API payload | | tools/convert_ic_lora.py | Diffusers IC LoRA → ComfyUI generic LoRA | | tools/build_workflow.py | Rebuild the draft-to-2K workflow | | assets/silence_audio_124f.safetensors | Official fixed-silence latent for the IC audio rows | | tests/test_nodes.py | Unit tests for preprocessing and sigma grids |

Nodes

Open-H3 IC · Prepare Draft (124f / 24fps)

  • samples 124 frames at 24 fps
  • cover-resizes and center-crops to 1344×768
  • trims or pads audio using its real sample rate
  • requires at least 113 real frames after resampling

Open-H3 IC · Official Prompt

Prepends the official enhancement task and preserves the source caption as Scene description.

Open-H3 IC · Sigma Grid

| Mode | Sigmas | |---|---| | up3 | 0.834942, 0.696852, 0.468753, 0 | | up4 | 0.923077, 0.834942, 0.696852, 0.468753, 0 | | hf8 | 1.0, 0.99391, 0.984288, 0.966064, 0.923077, 0.834942, 0.696852, 0.468753, 0 |

Open-H3 IC · Compose Target Latent

Combines the upscaled target video latent with the official 124-frame silence latent and pins the audio stream with noise_mask=0.

Installation

ComfyUI Manager

Search for Open-H3 Regenerate 2K IC in ComfyUI Manager, or install Registry ID comfyui-open-h3-regenerate-2k-ic.

Comfy Registry page

cd ComfyUI/custom_nodes
git clone https://github.com/slmonker/ComfyUI-OpenH3Regenerate2KIC.git

Restart ComfyUI after installation.

The T2VA example also uses:

  • ComfyUI-LynnReal — attention/feed-forward helpers
  • ComfyUI-Spark-H3 — Sol-Attn sparse attention
  • ComfyUI-VideoHelperSuite — VHS_VideoCombine

The draft-to-2K example uses LynnRealH3VAELoader for the fast H3 VAE decoder.

Common models

Place these under ComfyUI/models/:

models/
├── text_encoders/
│   └── MiniMax-H3/qwen3vl_32b_minimax_h3_int8_convrot.safetensors
├── vae/
│   └── MiniMax-H3/
│       ├── minimax_h3_video_vae_fp16.safetensors
│       └── sol_h3_vae.safetensors
├── latent_upscale_models/
│   └── minimax_h3_latent_upscaler_3d_conv_v1_bf16.safetensors
└── loras/
    ├── minimax_h3_hyperflow_8step_v1.0_comfyui_bf16.safetensors
    ├── minimax_h3_lms_v1.0_r64.safetensors
    └── minimax_h3_open_regenerate_2k_ic_comfy_bf16.safetensors

Download the converted IC LoRA from:

https://huggingface.co/slmonker/Open-H3-Regenerate-2K-IC-ComfyUI

Load the IC LoRAs in this order:

  1. HyperFlow 8-step
  2. LMS Sharp
  3. Open-H3 Regenerate 2K IC

T2VA-only models

models/
├── diffusion_models/
│   └── minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors/
│       └── minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors
├── vae/MiniMax-H3/
│   └── minimax_h3_audio_vae_fp32.safetensors
└── loras/
    └── dmad_minimax_h3_4step_full_critic_comfy.safetensors

The hybrid FL2VA+Ref2VA checkpoint and DMAD LoRA come from the MiniMax-H3 / TaoMate-HD ecosystem. The hybrid checkpoint is not included in this repository.

Running the T2VA example

  1. Open workflows/regen2k_ic_T2VA.json.
  2. Edit T2VA prompt · edit this text once.
  3. Run the graph.
  4. The 768P draft preview and 2K final video are written under ComfyUI/output/OpenH3/.

Default settings:

  • draft: 1344×768, 124 frames, 24 fps, DMAD 4-step
  • IC refinement: up3, CFG 1, 2560×1440

Running the draft-to-2K example

  1. Put an existing 768P draft into ComfyUI/input/.
  2. Open workflows/open_h3_regenerate_2k_ic.json.
  3. Select the draft in Load Video.
  4. Enter the original prompt in Open-H3 IC · Official Prompt.
  5. Run the graph.

For 1080P, set both resolution-bearing nodes to 1920×1088. For 4K use 3840×2176, which is usually impractical on a 32 GB GPU.

Convert the official IC LoRA

python tools/convert_ic_lora.py \
  open_h3_regenerate_2k_ic_lora.safetensors \
  minimax_h3_open_regenerate_2k_ic_comfy_bf16.safetensors \
  --dtype bf16

Conversion details:

  • q/k/v → fused qkv_proj with block-diagonal B
  • ff.net.0.proj from Diffusers [value;gate] to ComfyUI [gate;value]
  • fused qkv alpha = 3 × rank
  • output BF16

Released conversion:

  • 312 source linear layers
  • 208 ComfyUI modules
  • 624 tensors
  • zero missing base-model mappings

Validation

python -m unittest discover -s tests -v

The draft-to-2K workflow loads in ComfyUI and converts through graphToPrompt() with zero input-schema errors. The converted LoRA was checked against minimax_h3_ref2va_bf16.safetensors.

Notes

  • Default clip length is 124 frames (about 5.17 seconds).
  • The T2VA example keeps H3-generated draft audio in stage 1.
  • The IC stage uses fixed silence internally and outputs the draft audio.
  • 2K on 32 GB VRAM requires model offload / low-VRAM attention handling.
  • This is a community implementation, not an official MiniMax product.

Acknowledgements

License

Node code is Apache-2.0. MiniMax-H3 weights and derivative assets remain subject to the MiniMax-H3 Community License. See NOTICE.md.