Save 4 hours! We auto-setup your workflow! Free!

Drop your workflow.json — we handle every dependency, custom node, and model. Just open the link and run.

Auto-Setup Workflow Json (Free) Now!
ComfyUI > Nodes > Muse-MiniMax-H3-Unified-Loader

ComfyUI Extension: Muse-MiniMax-H3-Unified-Loader

Repo Name

Muse-MiniMax-H3-Unified-Loader

Author
muse-collective-26 (Account age: 0 days)
Nodes
View all nodes(1)
Latest Updated
2026-11-11
Github Stars
0.01K

How to Install Muse-MiniMax-H3-Unified-Loader

Install this extension via the ComfyUI Manager by searching for Muse-MiniMax-H3-Unified-Loader
  • 1. Click the Manager button in the main menu
  • 2. Select Custom Nodes Manager button
  • 3. Enter Muse-MiniMax-H3-Unified-Loader in the search bar
After installation, click the Restart button to restart ComfyUI. Then, manually refresh your browser to clear the cache and access the updated list of nodes.

Visit ComfyUI Online for ready-to-use ComfyUI environment

  • Free trial available
  • 16GB VRAM to 80GB VRAM GPU machines
  • 400+ preloaded models/nodes
  • Freedom to upload custom models/nodes
  • 200+ ready-to-run workflows
  • 100% private workspace with up to 200GB storage
  • Dedicated Support

Run ComfyUI Online

Muse-MiniMax-H3-Unified-Loader Description

Muse-MiniMax-H3-Unified-Loader is a ComfyUI node that loads H3 models, CLIP, VAEs, and LoRAs for Muse Minimax Director, with VRAM-aware profiles and optional attention and frame-model settings.

Muse-MiniMax-H3-Unified-Loader Introduction

Muse-MiniMax-H3-Unified-Loader is a ComfyUI custom node that prepares the model components used by Muse Minimax Director. Instead of connecting separate loaders for the main model, CLIP, VAEs, LoRAs, and attention settings, you can configure them in one place and send the resulting outputs to the Director.

The loader offers three configurable profiles: Low VRAM, Balanced, and Maximum Quality. You can choose one yourself or use Automatic to select a profile based on your GPU’s detected video memory (VRAM). This is useful if you work across different GPUs or want to compare model setups without rebuilding your workflow.

The extension does not include or download model weights. You choose the model and encoder files for each profile. It also does not change your prompts or modify the Director’s code.

How Muse-MiniMax-H3-Unified-Loader Works

Think of each profile as a saved equipment kit for a project. You decide which main model and CLIP encoder belong in each kit. When you run the workflow, the loader picks a kit, loads its components, applies your enabled options, and reports what it used.

In Automatic mode, the starting choice depends on detected VRAM:

Detected VRAMStarting profile
Up to 12.5 GBLow VRAM
More than 12.5 GB, up to 20.5 GBBalanced
More than 20.5 GBMaximum Quality

These names organize your file choices; they are not three model downloads supplied by the extension. If the preferred profile does not have a complete model-and-CLIP pair configured, the loader selects a nearby complete profile instead. It does not convert a model to a different format or quantization.

The loader routes H3 model loading through ComfyUI-H3-Multishot. Its adaptive VRAM system measures memory use for the render shape and refines its estimate over repeated runs. That helps manage memory more carefully, although it cannot guarantee that every model or render will fit on a given GPU.

To use the result, connect the loader’s model, clip, vae, and audio_vae outputs to the matching inputs on Muse Minimax Director. Connect model_fl2va as well if your Director workflow uses a First/Last-Frame model. The Director’s prompt_override remains a separate text input; this loader does not alter prompt text.

Muse-MiniMax-H3-Unified-Loader Features

Profiles and automatic selection

Set up a model and CLIP file for each profile you intend to use, then switch with the profile dropdown. Choose Automatic for hardware-based selection, or select Low VRAM, Balanced, or Maximum Quality directly when you want to test a particular setup.

Each profile has separate model, CLIP, and mmproj selections. Choose an mmproj file when your selected CLIP setup needs one. Leave a profile’s model or CLIP set to (disabled) if you do not want to make that profile available. Automatic selection can fall back to another complete profile.

Optional First/Last-Frame model

The model_fl2va slot loads a separate MiniMax H3 First/Last-Frame checkpoint for Director workflows that need it, such as Hybrid Continuation. If your workflow does not use that checkpoint, you can leave this slot unused.

Three LoRA slots

You can configure up to three LoRAs. Each slot has its own strength and target: the reference model, the First/Last-Frame model, or both. For example, you might apply one LoRA only to the reference model while applying another to both models. Changing a slot’s strength changes how strongly that LoRA is applied.

Attention and memory options

Advanced options include a ComfyUI core attention-backend selector, SageAttention, low-VRAM attention, and chunked feed-forward processing. Enabled patches are applied consistently to the reference and First/Last-Frame models.

The core attention-backend selector, SageAttention, and High-Res Stability default to Off. Low-VRAM attention and feed-forward chunking remain enabled by default. If you prefer to apply your own attention nodes after this loader, turn off the loader’s corresponding patches first, then place your nodes between its model or model_fl2va output and the Director.

SageAttention requires a compatible package and CUDA kernels for your environment. Leave it off unless you have confirmed that setup. The browser checks compatibility before queueing a run when SageAttention is enabled and displays an explanatory notice if it needs attention.

Run status

The status text output summarizes the profile, detected VRAM, LoRAs, and patches used for that run. Check it when Automatic selects a different profile than expected or when you want to confirm which optional settings were applied.

Muse-MiniMax-H3-Unified-Loader Models

The loader supports several model roles, but it does not bundle model files:

Model or componentWhat it is used for
Profile modelThe main H3 model selected by the active profile. Each profile can point to a different file.
CLIP and optional mmprojEncoder components selected alongside the profile model; choose an mmproj when your CLIP setup requires one.
model_fl2vaAn optional First/Last-Frame checkpoint for Director workflows that use that model.
VAEsComponents supplied through the vae and audio_vae outputs.
LoRAsOptional additions applied through three independently configurable slots.

The H3 loading path supports safetensors and GGUF model selections. A .gguf file also requires ComfyUI-GGUF. The extension does not prescribe a particular file for Low VRAM, Balanced, or Maximum Quality: the output and memory demands depend on the files and settings you choose.

What's New with Muse-MiniMax-H3-Unified-Loader

Version 1.3.0 — October 18, 2026

  • Updated ComfyUI compatibility: The author updated and render-tested the loader with ComfyUI 0.41.3, Python 3.12.10, and PyTorch 2.9.1+cu130 on Windows. This is a tested configuration, not a requirement to install that exact PyTorch/CUDA build. The RunPod/Linux package includes the same loader fixes, but this release does not claim a new Linux GPU render test.
  • Updated attention-backend support: The optional selector uses ComfyUI’s current Model Attention Backend execute API, with a legacy patch fallback. It sits under Advanced and defaults to Off.
  • More cautious defaults: SageAttention and High-Res Stability default to Off. Low-VRAM attention and feed-forward chunking remain enabled.
  • SageAttention checks: When enabled in the browser, compatibility is checked before queueing and an explanatory notice appears if setup needs attention. Direct API submissions also receive a defensive backend check.
  • Pinned-memory behavior: Default follows ComfyUI’s setting. Explicit caps do not exceed ComfyUI’s original limit; returning to Default restores that behavior, while disabled pinning stays disabled.
  • Improved Automatic fallback: When the preferred profile is disabled, Automatic chooses an available complete model-and-encoder profile. It does not download files or change their quantization.

Existing workflows retain their saved settings. After updating, restart ComfyUI and hard-refresh your browser so the updated node interface loads.

September 5, 2026

Missing-dependency errors were improved to name the required package and point to it. These dependencies are called internally, so ComfyUI Manager’s missing-node scan may not identify them from the saved workflow alone.

September 4, 2026

Sol-Attn was removed from the available attention options. In the author’s controlled GGUF Low VRAM test, enabling it caused a render to stall. Removing the option prevents you from selecting that known-problematic configuration through this node.

Troubleshooting Muse-MiniMax-H3-Unified-Loader

A run reports that a required node is missing

  1. Read the error message to identify the package.
  2. Check that ComfyUI-H3-Multishot is available. The loader requires its H3 loading nodes.
  3. If an enabled attention or memory patch needs it, check ComfyUI-KJNodes. Low-VRAM attention and feed-forward chunking are enabled by default.
  4. Restart ComfyUI so it can load the available nodes, then run the workflow again.

ComfyUI Manager may not offer an “install missing custom nodes” prompt for these packages because the loader calls their nodes internally rather than placing them separately in your workflow graph.

Automatic uses a different profile than expected

  1. Check the status output for detected VRAM and the profile actually selected.
  2. Check that your preferred profile has both a model and CLIP file selected, rather than (disabled).
  3. If you want to use one specific setup regardless of detected VRAM, select that profile directly in the profile dropdown.

A GGUF model does not load

Check that ComfyUI-GGUF is available. It is required when you select a .gguf model file. Restart ComfyUI after making it available, then retry the run.

SageAttention shows a compatibility notice or fails

  1. Turn SageAttention Off and retry to check whether it is the source of the problem.
  2. If you intend to use it, verify that its package and CUDA kernels match your environment before enabling it again.
  3. Keep SageAttention Off if you do not have a confirmed compatible setup; it is optional.

The interface still looks unchanged after an update

Restart ComfyUI, then hard-refresh the browser. Saved workflows keep their settings, so also check the node’s current toggles rather than assuming they changed to new defaults.

Frequently asked questions

Do I need all three profiles configured?
No. Configure the profiles you plan to use. If Automatic’s preferred profile is disabled, it can select another complete model-and-CLIP profile.

Does the loader supply or download MiniMax H3 weights?
No. You select the model files; weights are not included.

Does the loader edit my prompts or the Director node?
No. Prompt text stays with the Director’s inputs, and this loader supplies model components without modifying the Director’s code.

Can I use my own attention-patch nodes?
Yes. Turn off the corresponding patches in this loader, then connect your preferred patch node between the loader’s model output and the Director.

Learn More about Muse-MiniMax-H3-Unified-Loader

  • Muse Minimax Director: The companion workflow node and the place to check MiniMax H3 model licensing and attribution requirements.
  • ComfyUI-H3-Multishot: Background on the H3 loading and adaptive VRAM system used by this extension.
  • ComfyUI-GGUF: Relevant when your chosen model is a .gguf file.
  • ComfyUI-KJNodes: Relevant to the loader’s optional attention and memory patches.

The author licenses this extension’s original loader code under the MIT License. MiniMax H3 model weights are not part of that code; consult the Director repository for the model’s own licensing and attribution information.

Muse-MiniMax-H3-Unified-Loader Related Nodes

RunComfy
Copyright 2025 RunComfy. All Rights Reserved.

RunComfy is the premier ComfyUI platform, offering ComfyUI online environment and services, along with ComfyUI workflows featuring stunning visuals. RunComfy also provides AI Models, enabling artists to harness the latest AI tools to create incredible art.

Muse-MiniMax-H3-Unified-Loader detailed guide | ComfyUI