ComfyUI-OpenH3Regenerate2KIC Introduction
ComfyUI-OpenH3Regenerate2KIC provides ComfyUI helper nodes and example workflows for turning a 768P video draft into a higher-resolution, AI-refined video using Open-H3-Regenerate-2K/IC.
Rather than simply enlarging each frame, the workflow uses the draft as a visual reference, enlarges its compressed video representation, and applies a short AI refinement pass. The default output is 2560×1440, referred to as “2K” throughout the supplied workflows.
You can start with:
- A text prompt: Generate a 1344×768 draft, then refine it to 2560×1440.
- A reference image: Create a 768P reference-based video draft, then refine it to 2560×1440.
- An existing 768P draft: Skip draft generation and use the refinement workflow directly. This draft-first approach lets you develop the scene at a lower resolution before running the higher-resolution stage. It is useful when you already have a draft you want to enhance without rebuilding the entire generation workflow.
Important expectations:
- The default clip is 124 frames at 24 frames per second, or approximately 5.17 seconds.
- The final video uses the draft’s audio. The IC refinement stage does not generate replacement audio.
- High-resolution refinement remains demanding: 2K on a 32 GB GPU requires model offloading and low-VRAM attention handling.
- This is a community implementation, not an official MiniMax product. Project links: GitHub repository · ComfyUI-compatible IC LoRA · Upstream IC weights
How ComfyUI-OpenH3Regenerate2KIC Works
Think of the workflow as enlarging a rough painting and then giving it a finishing pass. The draft provides the scene and motion; the higher-resolution stage uses that draft and its description to guide refinement.
1. Prepare a consistent draft
For an existing video, the Prepare Draft node:
- Samples a 124-frame sequence at 24 fps.
- Resizes and center-crops the frames to 1344×768.
- Trims or pads the audio to match the prepared clip.
- Checks that there are at least 113 real frames after resampling. This gives the rest of the workflow a consistent input format.
“Cover-resize and center-crop” means the image is enlarged until it fills the required frame, then any excess around the edges is removed. A differently shaped source may therefore lose content near its borders.
2. Use the draft as a visual reference
The draft becomes <Video 1> in the H3 reference-to-video-and-audio stage, called Ref2VA.
In plain language, the model receives the draft as an example to follow rather than starting from text alone. The original scene description is also supplied through the Official Prompt node.
3. Enlarge the video in latent space
A latent is the model’s compressed working representation of the video—similar to a compact digital sketch that contains information needed to reconstruct the frames.
The workflow uses:
Draft video
→ H3 VAE encoding
→ Official 3D latent upscaler
→ Higher-resolution video latent
The VAE converts between visible frames and this compressed representation.
The learned latent upscaler enlarges the representation directly instead of using a conventional pixel upscaler. For an existing draft, the initial encoding is still necessary; the latent upscaling step avoids adding a separate decode → pixel-upscale → encode round-trip.
4. Apply a short refinement pass
Three LoRA adapters are loaded in this order:
- HyperFlow 8-step
- LMS Sharp
- Open-H3 Regenerate 2K IC A LoRA is an adapter that modifies how a larger model behaves; it is not a complete standalone model.
The IC stage uses a supplied sampling schedule, or sigma grid, to control its refinement steps. The default is up3, a three-step schedule.
Although the HyperFlow adapter has “8-step” in its name, the supplied IC workflow uses the schedule selected by the Sigma Grid node. The default IC refinement is therefore three steps, not eight.
5. Decode the result and restore the draft audio
The refined latent is decoded into visible frames. The final video is then combined with the draft’s audio.
Internally, the IC stage uses a fixed silence latent with its audio stream locked. This internal silence is not intended to replace the soundtrack in the exported video.
Pipeline overview
Text-to-video example:
Text prompt
→ MiniMax-H3 draft generation with DMAD 4-step LoRA
→ 1344×768 draft, 124 frames, 24 fps
→ Draft video and draft audio
→ Draft used as <Video 1> in Ref2VA
→ Official 3D latent upscaler
→ HyperFlow + LMS Sharp + Open-H3 IC LoRA
→ up3 refinement
→ 2560×1440 video with draft audio
Existing-draft example:
Existing 768P draft
→ Draft preparation and Ref2VA conditioning
→ H3 VAE encoding
→ Official 3D latent upscaler
→ IC refinement using up3, up4, or hf8
→ Fast H3 VAE decoding
→ Final video combined with the draft’s original audio
ComfyUI-OpenH3Regenerate2KIC Features
Three example workflows
| Starting point | Workflow | What it does |
|---|---|---|
| Text | regen2k_ic_T2VA.json | Generates a 1344×768 draft from your prompt, then refines it to 2560×1440. |
| Reference image | H3_regen2k_ic_Ref2va.json | Creates a 768P Ref2VA draft from a reference image, then refines it to 2560×1440. |
| Existing draft | open_h3_regenerate_2k_ic.json | Refines an existing 768P draft to 2560×1440. |
Four helper nodes
Open-H3 IC · Prepare Draft (124f / 24fps)
Standardizes the source video and audio before refinement.
- Frame rate: 24 fps.
- Prepared length: 124 frames.
- Prepared size: 1344×768.
- Minimum source coverage: 113 real frames after resampling.
- Audio handling: Trims or pads using the audio’s actual sample rate. For best control over composition, use a draft whose framing is already close to the example’s landscape format.
Open-H3 IC · Official Prompt
Adds the official enhancement instruction before your caption and preserves your caption under Scene description.
You supply the scene description; the node supplies the enhancement-task wording.
For example, if the draft shows a fox walking through snow, use the original description of that scene rather than an unrelated prompt. This keeps the text guidance aligned with the visual reference.
Open-H3 IC · Sigma Grid
Selects the sequence of noise levels used during refinement.
| Mode | Refinement steps | Sigma values |
|---|---|---|
up3 | 3 | 0.834942, 0.696852, 0.468753, 0 |
up4 | 4 | 0.923077, 0.834942, 0.696852, 0.468753, 0 |
hf8 | 8 | 1.0, 0.99391, 0.984288, 0.966064, 0.923077, 0.834942, 0.696852, 0.468753, 0 |
Start with up3, the supplied default. You can compare up4 or hf8 on the same draft when exploring alternatives.
More steps add sampling work, but the provided documentation does not establish that one mode always produces a better artistic result.
Open-H3 IC · Compose Target Latent
Combines:
- The upscaled target video latent.
- The official 124-frame silence latent.
It locks the internal audio stream with
noise_mask=0, meaning that stream is not changed by sampling. The workflow separately retains the draft audio for the final output.
Shared prompt editing in the text-to-video workflow
The T2VA example uses one text node for both stages:
- The draft generator receives your raw prompt.
- The IC stage receives the enhancement instruction plus the same prompt. You only need to edit “T2VA prompt · edit this text” once, reducing the chance of mismatched descriptions between draft generation and refinement.
Adjustable output resolution
The supplied draft-to-2K workflow documents these targets:
| Target | Dimensions to use |
|---|---|
| Default 2K | 2560×1440 |
| 1080P-oriented target | 1920×1088 |
| 4K-oriented target | 3840×2176 |
Change both resolution-bearing nodes together. Leaving one at the previous size can create inconsistent target dimensions.
The 1080P- and 4K-oriented settings use model-friendly aligned heights, not exactly 1080 or 2160 pixels. The author notes that the 4K setting is usually impractical on a 32 GB GPU.
ComfyUI-OpenH3Regenerate2KIC Models
The extension is a workflow integration, not a standalone video model. It relies on MiniMax-H3 components, an H3 latent upscaler, and several adapters.
Shared components
| Model file | Role |
|---|---|
qwen3vl_32b_minimax_h3_int8_convrot.safetensors | Text encoder that converts your prompt into information the H3 model can use. |
minimax_h3_video_vae_fp16.safetensors | H3 video VAE for converting between frames and video latents. |
sol_h3_vae.safetensors | Fast H3 VAE decoding component used by the supplied workflows. |
minimax_h3_latent_upscaler_3d_conv_v1_bf16.safetensors | Learned 3D upscaler that increases the spatial resolution of H3 video latents. |
minimax_h3_hyperflow_8step_v1.0_comfyui_bf16.safetensors | ComfyUI-compatible HyperFlow adapter used in the refinement stack. |
minimax_h3_lms_v1.0_r64.safetensors | LMS Sharp adapter used in the refinement stack. |
minimax_h3_open_regenerate_2k_ic_comfy_bf16.safetensors | ComfyUI-converted Open-H3 IC adapter for the regeneration stage. |
Load the refinement adapters in the documented order: HyperFlow → LMS Sharp → Open-H3 IC.
The converted IC adapter is available from the ComfyUI IC LoRA model repository. The author checked this conversion against minimax_h3_ref2va_bf16.safetensors.
Additional components documented for T2VA
| Model file | Role |
|---|---|
minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors | Hybrid FL2VA+Ref2VA checkpoint used by the T2VA example. |
minimax_h3_audio_vae_fp32.safetensors | Audio VAE used for draft-stage audio. |
dmad_minimax_h3_4step_full_critic_comfy.safetensors | Four-step DMAD adapter used to generate the initial draft. |
The hybrid checkpoint and DMAD adapter come from the MiniMax-H3 / TaoMate-HD ecosystem. The hybrid checkpoint is not included in the extension repository.
Understanding the upscaler choice
The supplied Open-H3 IC workflows use the 3D latent upscaler.
The separate H3 Latent Upscaler project also documents a newer 2D model with a smaller footprint and adjustable sharpness. Those are features of that separate project, not documented options in the supplied Open-H3 IC workflows.
Do not assume that its 2D checkpoint can replace the 3D checkpoint without workflow changes. The upstream upscaler requires the node architecture and checkpoint architecture to match.
Using ComfyUI-OpenH3Regenerate2KIC
Generate from text
- Open
workflows/regen2k_ic_T2VA.json. - Find “T2VA prompt · edit this text”.
- Enter your scene description.
- Run the workflow.
- Review the draft preview and final video under
ComfyUI/output/OpenH3/. Default settings:
| Stage | Settings |
|---|---|
| Draft generation | 1344×768, 124 frames, 24 fps, DMAD four-step |
| IC refinement | up3, CFG 1, 2560×1440 |
CFG controls text guidance during sampling. Start with the supplied value of 1 when evaluating the example.
Refine an existing draft
- Put your draft video in
ComfyUI/input/. - Open
workflows/open_h3_regenerate_2k_ic.json. - Select the video in Load Video.
- Enter the original scene prompt in Open-H3 IC · Official Prompt.
- Keep
up3and the default target size for your first test. - Run the workflow and compare the result with the source.
Start from a reference image
Open workflows/H3_regen2k_ic_Ref2va.json for the reference-image example. Its documented sequence is reference image → 768P Ref2VA draft → 2560×1440 refinement.
ComfyUI-OpenH3Regenerate2KIC Demo Comparisons
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: 768P draft on the left and 2K IC result on the right](https://github.com/slmonker/ComfyUI-OpenH3Regenerate2KIC/releases/download/demo-videos/Regen2KIC_00002-audio.mp4)
Demo 2
[Demo 2: 768P draft on the left and 2K IC result on the right](https://github.com/slmonker/ComfyUI-OpenH3Regenerate2KIC/releases/download/demo-videos/Regen2KIC_00003-audio.mp4)
Demo 3
[Demo 3: 768P draft on the left and 2K IC result on the right](https://github.com/slmonker/ComfyUI-OpenH3Regenerate2KIC/releases/download/demo-videos/Regen2KIC_00004-audio.mp4)
Troubleshooting ComfyUI-OpenH3Regenerate2KIC
The draft is rejected because it has too few frames
The preparation node requires at least 113 real frames after resampling to 24 fps.
- Check the source clip’s duration and frame rate.
- Remember that the original frame count may change when resampled to 24 fps.
- Use a longer source segment if the resampled clip falls below the minimum.
- Run the preparation stage again. The default output length remains 124 frames; that does not mean an arbitrarily short input is supported.
The framing is cropped unexpectedly
The draft preparation node fills 1344×768 and center-crops the excess.
- Compare the source’s aspect ratio with the target frame.
- Check whether important subjects sit close to the source edges.
- Reframe your source before loading it if those edges must remain visible.
The workflow runs out of GPU memory
- Check the target resolution. Avoid the documented 4K setting on a 32 GB GPU.
- Use model offloading and low-VRAM attention handling; the author identifies these as necessary for 2K on 32 GB.
- If needed, try the documented lower target of 1920×1088.
- Update both resolution-bearing nodes before running again. Latent upscaling can save processing time, but the refinement pass still operates at the target resolution. It does not remove the high-resolution memory requirement.
ComfyUI reports missing nodes
The examples depend on nodes beyond the four helpers in this repository.
Check which node names are missing against these documented dependencies:
| Workflow/component | Dependency |
|---|---|
| T2VA attention/feed-forward helpers | ComfyUI-LynnReal |
| T2VA Sol-Attn sparse attention | ComfyUI-Spark-H3 |
| T2VA video combination | ComfyUI-VideoHelperSuite, including VHS_VideoCombine |
| Draft-to-2K fast VAE decoding | LynnRealH3VAELoader |
Identify the missing component before changing the workflow’s connections or replacing its nodes.
A model is missing or an adapter fails to load
- Compare the selected filename with the model tables above.
- For IC, confirm that you selected the ComfyUI-converted adapter rather than the upstream Diffusers-format file.
- Check the adapter order: HyperFlow → LMS Sharp → Open-H3 IC.
- For T2VA, check the separate hybrid checkpoint, audio VAE, and DMAD adapter.
- For the latent upscaler, confirm that the selected 3D node and 3D checkpoint match.
The final video is silent or has the wrong audio
- Confirm that the draft itself contains the intended soundtrack.
- Check that the prepared draft audio reaches the final video-combination stage.
- Do not use the IC stage’s internal fixed-silence stream as the final soundtrack. The intended output combines refined video with draft audio; refinement does not create a new soundtrack.
Frequently asked questions
Is this a conventional video upscaler? No. It combines learned latent upscaling with AI refinement guided by the draft and its scene description.
Can I refine a draft without generating it again? Yes. Use the draft-to-2K workflow.
Does the extension generate audio at the refinement stage? No. The T2VA draft stage can generate audio, and that audio is retained for the final video. The IC stage uses fixed silence internally.
Does hf8 guarantee a better result than up3?
No such guarantee is documented. Compare the modes on your own draft; up3 is the supplied default.
Can I process arbitrary clip lengths? The supplied workflow is built around 124 frames at 24 fps. Arbitrary-duration processing is not documented.
Are HyperFlow’s published benchmark times the expected speed of this extension? No. Those measurements describe separate upstream HyperFlow pipelines and hardware. They are not benchmarks for this two-stage ComfyUI workflow.
Learn More about ComfyUI-OpenH3Regenerate2KIC
- Extension repository — The author’s workflows, node code, and project documentation.
- Open-H3-Regenerate-2K/IC upstream documentation — Background on the refinement method integrated here.
- ComfyUI-compatible IC LoRA — Converted adapter used by this extension.
- Upstream IC model repository — Original model assets.
- MiniMax-H3 and Comfy-Org MiniMax-H3 — Base-model information and ComfyUI-oriented resources.
- HyperFlow documentation — Explanation of the acceleration adapter. Its command-line and Diffusers instructions describe the upstream project, not this extension’s ComfyUI controls.
- H3 Latent Upscaler documentation — Further explanation of learned latent upscaling, resolution alignment, and memory considerations.
ComfyUI-OpenH3Regenerate2KIC License
The extension’s node code is licensed under Apache-2.0.
MiniMax-H3 weights and derivative assets have separate licensing requirements under the MiniMax-H3 Community License. The code license does not replace the model licenses.
Review NOTICE.md and the applicable model licenses before using or distributing generated work, models, or derivative assets.
