COMFYUI / LOCAL WORKFLOW GUIDE
MiniMax H3.1 ComfyUI Setup and Workflow
MiniMax H3.1 ComfyUI setup currently means using the available MiniMax H3 weights and native ComfyUI workflow. Follow the official text-to-video template below to select model files, import the graph, and prepare your first local render. H3.1 availability is covered separately; this guide does not rename H3 weights or present an unannounced model as downloadable.
For a browser-based creative brief, return to the MiniMax H3.1 video generator and work out your prompt before opening ComfyUI.
OFFICIAL COMFYUI EXAMPLE
A practical guide with official download links. Models run in your own ComfyUI installation; this website does not host weights or execute the workflow.
Requirements
Use a current ComfyUI release before selecting files. This walkthrough follows the native T2V template and its INT8 checkpoint, rather than a third-party wrapper. Keep the application, its bundled dependencies, and the selected quantization compatible with one another.
| Requirement | What to prepare | Official source |
|---|---|---|
| ComfyUI version | Update to the current release. Basic H3 workflows started at 0.30.0; current template metadata names 0.33.0, and INT8 video-VAE performance support requires 0.36.0 or later. | Official source |
| VRAM and system memory | Official details not yet announced. There is no single published minimum for this exact native template. Capacity depends on precision, offloading, output size, and duration; do not treat a community GPU example as a guaranteed minimum. | Official source |
| Operating system | This INT8 walkthrough targets Windows or Linux with a compatible NVIDIA setup. General ComfyUI platform support does not guarantee that every H3 quantization backend works on every device. | Official source |
| Dependencies | Use the matching ComfyUI distribution. The model card recommends PyTorch cu130 for INT8 convrot. Core requirements include torch, transformers, safetensors, av, and comfy-kitchen; avoid independently mixing package versions. | Official source |
| Model-file size | The five selected files total 42,030,035,159 bytes: approximately 42.03 GB, or 39.14 GiB. This is disk storage for your local installation, not a VRAM requirement. Leave additional room for ComfyUI, temporary downloads, and outputs. | Official source |
Start with the template’s small draft resolution and a short clip. A successful small render is a useful compatibility check, but it does not establish that a longer or larger job will fit the same hardware.
Download the weights
Use Comfy-Org’s repackaged MiniMax-H3 files for this graph. The original MiniMax release uses a different repository structure and is not a substitute for these loader selections. Keep filenames intact and place each file in its matching model directory, rather than putting everything in checkpoints.
Route 1 · Hugging Face
Open Comfy-Org/MiniMax-H3, or use the individual file links in the table. Download all five selections. Compare the completed download size with the file listing; a small Git LFS pointer or an HTML download page is not a model.
Open Hugging Face filesRoute 2 · ModelScope
Open Comfy-Org/MiniMax-H3 on ModelScope, enter Files and versions, and browse the same folders and filenames shown below. Use that repository’s download controls. This is an alternative source for the ComfyUI package, not instructions to rename the original MiniMax files.
Open ModelScope files| File and role | Place inside your ComfyUI folder | Download size | Official file |
|---|---|---|---|
Diffusion modelminimax_h3_fl2va_pruned_int8_convrot.safetensors | models/diffusion_models/ | 20.97 GB | Download fileFile details |
Text encoderqwen3vl_32b_minimax_h3_nvfp4_awq.safetensors | models/text_encoders/ | 15.69 GB | Download fileFile details |
Video VAEminimax_h3_video_vae_int8_convrot.safetensors | models/vae/ | 2.81 GB | Download fileFile details |
Audio VAEminimax_h3_audio_vae_fp32.safetensors | models/vae/ | 605.25 MB | Download fileFile details |
Template’s turbo LoRAminimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors | models/loras/ | 1.96 GB | Download fileFile details |
The LoRA is included in the template’s model scan even when turbo mode is disabled. Optional style embeddings are not required for the first run. Sizes above use decimal units from the Hugging Face file metadata; we have not independently established byte-for-byte identity of every ModelScope mirror file.
Install nodes
The native H3 graph uses ComfyUI core nodes. Manager can help inspect an imported workflow and manage genuinely additional packages, but installing a similarly named community extension does not replace updating the core application.
Update the installation you actually launch.
Follow ComfyUI’s official update guide for your installation type. If you keep several installations, check which directory your launcher starts; updating an unused copy will not change the nodes in the open browser tab.
Official sourceOpen Manager if it is available.
Desktop or your chosen distribution may already include it. If it is absent, use the official Manager installation guide for that distribution. There is no extra H3-specific custom-node package to install for this native template.
Official sourceImport the official JSON and inspect missing nodes.
Use Manager’s missing-node tools where applicable. If MiniMaxH3ImageToVideo is reported missing, update ComfyUI core and inspect startup import failures before searching for an unrelated custom-node pack.
Official sourceRestart and refresh after updates.
Restart the running ComfyUI service, refresh its browser interface, and import the JSON again. Install an additional community package only if a modified workflow actually names that package and you have checked its author and requirements.
Official source
Import the workflow
Download video_minimax_h3_t2v.json from Comfy-Org, then drag the JSON into the ComfyUI canvas or use Workflow → Open. The first/last-frame inputs stay empty for text-to-video. Select the five model filenames in their loaders, enter a short prompt, and keep turbo disabled for your initial compatibility check.
Keep ResolutionSelector connected to the H3 subgraph’s width and height. The template uses 16:9, 0.4 megapixels, and a multiple of 32 for a small draft.
Open the subgraph when checking connections. CLIPLoader uses type minimax; the video VAE feeds both the H3 node and video decoder. Positive conditioning goes to BasicGuider and the audiovisual latent goes to SamplerCustomAdvanced.
Keep both decode branches: video through VAEDecode, audio through VAEDecodeAudio with the separate audio VAE. Feed both into CreateVideo, then connect its VIDEO output to SaveVideo.
For a red box, identify its exact node type. A missing core H3 node calls for an application update; a loader with an empty model choice calls for the correct file and directory. Do not replace every red item with a random node from Manager.
Run a small job and inspect the saved video and audio before changing multiple settings. Keep a copy of the unmodified JSON so you can compare connections after experimenting.
The download points to the upstream workflow, which may change after this review date. If its filenames differ, follow the current template and its model links together instead of mixing parts from different revisions.
Key parameters
These are starting settings for the available H3 native graph. They are not H3.1 specifications or measurements from this website. Change one setting at a time so you can associate a failure or quality change with a specific choice.
| Parameter | Suggested starting point or range | Effect and boundary |
|---|---|---|
| Steps | 20 for the base workflow; try 20–25. Use 8 only with the matching turbo LoRA and turbo mode. | More sampling work costs time. Switching to a turbo step count without its matched configuration does not reproduce the turbo workflow. |
| CFG | Not exposed in this native template. Keep BasicGuider. | There is no CFG slider to set to a generic range. A custom CFG graph is a separate configuration, not a required installation step. |
| Resolution | Draft: 0.4 MP at 16:9, approximately 864×480. Native target: 0.98 MP, approximately 1344×768; dimensions aligned to 32. | Higher area increases resource use. Keep within the native canvas for this workflow; 2K delivery needs a separate upscale stage. |
| Duration | Start with 5 requested seconds. Stay around the documented 5–15-second training range when exploring. | The graph aligns length to 17k+5 frames. Five requested seconds becomes 124 frames, about 5.17 seconds at 24 fps. Local node acceptance is not the hosted API’s duration policy. |
| Frame rate | Keep 24 fps. | Keep the duration calculation and CreateVideo timeline consistent. Changing only the save frame rate changes playback speed rather than creating new model motion. |
| Sampler and scheduler | Keep res_multistep and simple from the template. | Establish the baseline first. Replace attention or sampling components separately so that a regression has a useful comparison. |
Troubleshooting
The entries below describe real failure categories and checks, not invented H3 error codes. Capture the failing node name and the first useful terminal message. Restore the official baseline before adding optimizations or changing several files at once.
| Symptom | What to check | Next action |
|---|---|---|
| GPU out of memory | Resolution, clip length, checkpoint precision, and offloading all affect memory. | Return to the small draft and short clip; close competing GPU jobs. If that still fails, review the model card’s supported precision path for your hardware rather than assuming a universal VRAM minimum.Diagnosis source |
| Missing Node Packs / red H3 node | The official issue names MiniMaxH3ImageToVideo as missing on an outdated installation. | Update the ComfyUI core you run, restart it, and reimport. A similarly named custom package is not a substitute for the native node.Diagnosis source |
| Black or corrupted frames | Checkpoint/VAE compatibility, incomplete downloads, decoder wiring, and attention backend changes. | Restore the template’s video VAE and default attention path, then test a draft. The official guide documents incompatible INT8 attention combinations; change optimizations one at a time.Diagnosis source |
| Broken motion or unexpected frame counts | Seconds may have been entered as frames, or the duration-expression connection may have changed. | Restore the duration-to-length connection, 17k+5 frame alignment, and 24 fps timeline. Compare the actual output frame count before diagnosing playback artifacts as a model failure.Diagnosis source |
| Video has no audio | The separate audio VAE, VAEDecodeAudio, and CreateVideo.audio branch. | Reconnect decoded audio to CreateVideo before SaveVideo. Saving images alone does not save a soundtrack; also check that the player is unmuted.Diagnosis source |
| Clip is too long or the job stalls | The requested duration, resulting frame count, available memory, and training range. | Return to a short request. The local node can accept lengths beyond the usual training range; do not expect a made-up 15-second error code or assume that acceptance guarantees a usable result.Diagnosis source |
| LoRA shape mismatch | A full-build LoRA may not match the pruned checkpoint’s tensors. | Match the LoRA to its checkpoint build. Keep the supplied turbo combination for the first run; renaming a file cannot make incompatible weights compatible.Diagnosis source |
H3.1 in ComfyUI
Official details not yet announced.
After an official H3.1 release, this section will be updated with verified weight links, node compatibility, and a corresponding workflow. Until then, the working instructions above remain explicitly about H3.
MiniMax official announcementsFAQ
Can I use this MiniMax H3.1 ComfyUI guide today?
Yes, the practical setup sections use the available H3 weights and the official ComfyUI text-to-video template. The H3.1 name is the subject of the availability section, not a downloadable checkpoint. Follow the linked files and graph together, and begin with a small local compatibility check before planning larger renders.
Do I need to install a separate MiniMax H3 custom node?
The native template uses ComfyUI core nodes, including MiniMaxH3ImageToVideo. Start by updating the installation you actually launch. Manager is useful for diagnosing additional packages, but an unrelated extension does not replace a missing core implementation. A modified community workflow may have different requirements, so inspect its own node list separately.
Can I download the weights from ModelScope instead of Hugging Face?
Yes. The Comfy-Org MiniMax-H3 repository is available on both services. On ModelScope, browse the same directories and filenames listed in this guide. Do not substitute the original MiniMax repository layout for the repackaged loader files. Check completed downloads against their own file listings before selecting them in your local workflow.
Does this website download or run the model for me?
No. This page provides instructions and official external links. The model files belong in your own ComfyUI installation, and local rendering uses your own hardware and storage. Visiting the tutorial does not download the weight pack or start a generation job. The homepage separately provides a browser-based brief preparation interface.
Is 42 GB the amount of GPU memory I need?
No. That figure describes the combined disk size of the selected model files in decimal units. Runtime GPU and system memory depend on loading, offloading, quantization, resolution, and clip length. We have not published a measured hardware matrix for this exact setup, so the file total should not be treated as a VRAM recommendation.
What CFG value should I use for the official H3 workflow?
The linked native graph uses BasicGuider and does not expose a CFG control. Leave that arrangement intact for the baseline run. A generic CFG range from another video model is not an instruction for this template. If you deliberately build a different guidance graph, document and test that configuration separately.
Why does a five-second request produce more than 120 frames?
The workflow converts seconds into a frame count aligned to H3’s seventeen-frame block grid with a five-frame offset. A five-second request therefore becomes 124 frames rather than 120. Keep that calculation connected to the model length and preserve the matching playback rate instead of manually trimming intermediate latents.
Can I switch the native workflow directly to 2K?
Treat native model generation and final delivery resolution as separate stages. The ComfyUI guide describes a native canvas around a 768-pixel short edge, with a larger output produced through a separate upscale pass. Confirm a working base render first. The hosted service’s resolution menu is not a local workflow setting.
Before refining a local workflow, prepare a clear subject, action, camera move, and sound direction in the MiniMax H3.1 video generator and bring that brief back to your own ComfyUI canvas.