Open H3 Regenerate 2K IC in ComfyUI: Sharper MiniMax H3 Video That Keeps Faces

The first open 2K refiner for MiniMax H3 made drafts sharper, but faces, mouth shapes and small text tended to drift, because it re-noised the draft and repainted it. NVlabs has now released Open-H3-Regenerate-2K-IC, an “in-context” version that keeps your clean 768P draft in view while it draws the high-resolution clip next to it.

The result follows the draft much more closely: the same eyes, the same lip sync, and signs that still read correctly.

This guide covers what changed, every file and folder for the community ComfyUI node pack, the three sampler modes, settings for 1080p, 2K and 4K, VRAM on a 32 GB card, and fixes for common errors.

New to H3? Our MiniMax H3 ComfyUI guide lists every H3 workflow, file and fix in one place.

What the IC Version Changes

NVlabs released the first Open-H3-Regenerate-2K in late October, and we covered it in our H3 2K and 4K upscale guide. It works by adding noise back into the upscaled draft and then cleaning it up at high resolution.

That adds lovely texture, but the model is free to repaint details, so faces, mouth shapes and small lettering can shift.

The IC version takes a different route:

• The clean draft stays in the transformer’s context as a reference video, and the high-resolution clip is drawn beside it.
• Because the draft is always visible, identity, lip sync, sign text and layout follow it.
• The trade-off: it adds less new texture than the first version, but keeps eyes, wrinkles and mouth shapes where the draft put them.

The project’s own car-park comparison shows the difference. The official H3-Regenerate-2K API turns the sign into unreadable marks, while the IC version still reads “PRIVATE CAR PARK”.

NVlabs does not spell out what IC stands for, but the release describes it as an in-context version, so in-context is the obvious reading.

How In-Context Regeneration Works

H3’s Ref2VA transformer can read reference videos as extra tokens. The IC LoRA is trained on top of that frozen transformer:

1. Each training pair is a real 768P-to-4K clip.
2. A downscaled copy of the clip is the clean reference.
3. The 4K target is noised by a random amount, from pure noise to almost clean.
4. The model learns to produce the target while looking at the reference, and the loss only counts the target tokens.

At run time, your 768P draft plays the part of the clean reference, and the model draws the 2K version alongside it. The audio is copied straight from the draft, untouched.

Two Ways to Run It in ComfyUI

There are no official ComfyUI nodes yet. Two community routes exist:

The official FP32 LoRA is about 8.3 GB, so both community files are smaller. The rest of this guide follows the slmonker pack, because it ships the workflows and the helper nodes that set up the draft, prompt and sigmas correctly.

Files and Folders

All paths are inside ComfyUI/models/. These are the files the slmonker pack expects for the draft-to-2K workflow.

Load the three LoRAs in this order: HyperFlow 8-step, then LMS Sharp, then the Regenerate 2K IC LoRA.

You also need the H3 Ref2VA model itself. The base files are on Comfy-Org/MiniMax-H3. The latent upscaler comes from LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler, and HyperFlow from Video-Rebirth/hyperflow. Our HyperFlow H3 Flash guide covers HyperFlow in more detail.

The text-to-video example needs three more files:

• diffusion_models/: minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors (not included in the repo; check where its folder lands after download, because the README shows it nested in a folder of the same name).
• vae/MiniMax-H3/: minimax_h3_audio_vae_fp32.safetensors.
• loras/: dmad_minimax_h3_4step_full_critic_comfy.safetensors.

Install the Node Pack

1. Open a terminal in ComfyUI/custom_nodes.
2. Clone the pack: git clone https://github.com/slmonker/ComfyUI-OpenH3Regenerate2KIC.git
3. For the draft-to-2K workflow, also install ComfyUI-LynnReal, which provides the LynnRealH3VAELoader fast VAE loader.
4. For the text-to-video example, also install ComfyUI-Spark-H3 (sparse attention) and ComfyUI-VideoHelperSuite (for VHS_VideoCombine).
5. Restart ComfyUI.

The easiest way to get the extra packs is to open the workflow and use ComfyUI Manager‘s “Install Missing Custom Nodes”.

The pack adds four nodes:

• Open-H3 IC · Prepare Draft (124f / 24fps): samples 124 frames at 24 fps, resizes and centre-crops to 1344×768, and trims or pads the audio. It needs at least 113 real frames after resampling.
• Open-H3 IC · Official Prompt: adds the official enhancement task text and keeps your caption as the scene description.
• Open-H3 IC · Sigma Grid: sets the sigma schedule for up3, up4 or hf8.
• Open-H3 IC · Compose Target Latent: joins the upscaled video latent with the official silent audio latent and keeps the audio stream untouched.

Workflow: Draft to 2K

Use this when you already have an H3 clip at 768P.

1. Copy your 768P draft video into ComfyUI/input/.
2. Open workflows/open_h3_regenerate_2k_ic.json from the pack.
3. Pick the draft in the Load Video node.
4. Paste the original prompt that made the draft into Open-H3 IC · Official Prompt. NVlabs says to pass the caption unchanged.
5. Queue the run.

The output is 2560×1440 by default. Drafts made at 384P, 512P or 576P come back at 768P instead.

There is also an API-format version, open_h3_regenerate_2k_ic_api.json, if you drive ComfyUI from scripts.

Workflow: Text to Draft to 2K

The examples/T2VA/regen2k_ic_T2VA.json workflow does everything in one go: it makes a 1344×768 draft from your prompt and then regenerates it at 2K.

1. Open the workflow.
2. Edit the node titled T2VA prompt · edit this text.
3. Queue it. Results land in ComfyUI/output/OpenH3/.

Its defaults are a 124-frame, 24 fps draft with the DMAD 4-step LoRA, then IC mode up3 at CFG 1, with a final output of 2560×1440. If you write your own H3 prompts, our H3 prompt guide helps.

Modes, Resolutions and Settings

All three modes use CFG 1. The LMS Sharp strength defaults to 1.0, and 0 turns it off.

The 1080p and 4K sizes end in 1088 and 2176 rather than 1080 and 2160, because the model needs dimensions that divide evenly. The README says to change both resolution-bearing nodes when you switch.

VRAM and Hardware

Neither NVlabs nor the node author gives exact VRAM numbers. What the node README does say:

• 2K on a 32 GB GPU needs model offload or low-VRAM attention.
• 4K at 3840×2176 is “usually impractical” on a 32 GB GPU.

The official inference code can split the transformer across 1, 2, 4, 7 or 8 GPUs, and its example command uses 8, which tells you how heavy the full-quality path is.

So on a single consumer card, start at 1080p with up3, then try 2K with offload. On 16 GB, follow every tip in our H3 out of memory guide and expect long runs. If you only need a bigger file rather than a regenerated one, Tessera‘s tiled upscaling is lighter.

IC vs First Version vs Other H3 Upscalers

Pick IC when the clip has a talking face or readable text. Pick the first version when you want the most added texture on scenery and do not mind small changes.

Licence

The weights are under the MiniMax H3 Community License. The model page says it excludes the EU, UK, South Korea and the US, and third-party parts keep their own licences. It also notes that Open-H3-Regenerate-2K-IC is not an official MiniMax product. Read the licence before any commercial use.

Troubleshooting

Prepare Draft fails or complains about frames

The node needs at least 113 real frames after resampling to 24 fps. Use a longer draft, or make sure your clip is not being cut short by the Load Video frame limit.

The output looks like the first version, with drifting faces

Check the LoRA order: HyperFlow 8-step, LMS Sharp, then the IC LoRA. Make sure the IC LoRA is applied to the Ref2VA model, not FL2VA.

LynnRealH3VAELoader or VHS_VideoCombine is red

Install ComfyUI-LynnReal (needed for the fast VAE) and VideoHelperSuite (for the T2VA example), then restart.

Out of memory at 2K

Turn on model offload or a low-VRAM attention mode, drop to 1080p, and close other GPU apps. 4K on 32 GB is not expected to work well.

Size mismatch errors after changing resolution

Both resolution-bearing nodes must match. Use 1920×1088 for 1080p and 3840×2176 for 4K, not 1080 or 2160.

The latent upscaler model is not found

It goes in ComfyUI/models/latent_upscale_models/, not upscale_models.

FAQ

What does IC stand for?

NVlabs calls it an in-context version and does not expand the letters. It keeps the draft in context while generating.

Does it change the audio?

No. The audio is copied unchanged from the draft. A silent audio latent ships with the weights for the internal steps.

Which file should I use, Kijai’s or slmonker’s?

slmonker’s full-rank file plus its workflows if you want the official recipe. Kijai’s 200 MB file if you want to drop the LoRA into your own Ref2VA graph.

Can I use it on non-H3 videos?

It is trained to regenerate H3 drafts and runs on the H3 Ref2VA model. The Prepare Draft node will resize other clips to 1344×768, but results on non-H3 footage are not documented.

Do I need the original prompt?

NVlabs recommends passing the caption that generated the draft, unchanged. The reference code also accepts an empty caption.

Is it faster than the first version?

The default up3 mode is three transformer calls, so it is quick for a refiner. No timing figures are published.

Sources

• Efficient-Large-Model/Open-H3-Regenerate-2K-IC (model card, settings, licence)
• slmonker/ComfyUI-OpenH3Regenerate2KIC (nodes, files, workflows, VRAM notes)
• ComfyUI Wiki: Open-H3-Regenerate-2K-IC
• Kijai/MiniMax-H3-experimental