top of page
MiniMax H3 All-In-One ComfyUI One-Click Windows Installer

MiniMax H3 All-In-One ComfyUI One-Click Windows Installer

Generate video with native stereo audio in a single pass — no separate audio step. This all-in-one installer sets up ComfyUI Portable, every custom node, and the models for all six MiniMax H3 workflows in one double-click, with a download menu so you pick exactly which models to pull instead of being forced into 66 GB.

Six ready-to-run workflows: FL2V (standard first/last-frame image-to-video and text-to-video), Ref2va (reference-image-to-video, better at holding identity and composition), LightX2V Turbo and Larryvrh Turbo (4–8 step generation), plus two lip-sync workflows that drive a portrait to sing or speak from a reference audio track.

What the installer does: sets up ComfyUI Windows Portable with the CUDA 12.8 PyTorch build, installs all eight required custom node packs (KJNodes, rgthree, Nvidia RTX Nodes, Memory Cleanup, Fearnworks, Spectrum-MiniMax-H3, Larryvrh H3-Turbo, VideoHelperSuite), then pauses and asks which models you want before downloading up to seven files into the correct folders. Git, 7-Zip and the right PyTorch build are all handled automatically.

Pick your footprint: Everything (~66 GB), Standard only (~63 GB), Turbo only (~45 GB), Ref2va only (~44 GB — the leanest complete path for lip sync and image-to-video), or Skip and fetch the models yourself later. Downloads are resumable, so a dropped transfer picks up where it left off rather than restarting.

Doing this by hand means a portable ComfyUI, a pinned CUDA build of PyTorch, eight node packs and up to seven model files across several Hugging Face repos. This installer does all of it.

  • What you get

    A one-click Windows installer that sets up ComfyUI Portable, all eight custom node packs, and every MiniMax H3 model — with an interactive menu that lets you choose your download size.

    Six workflows included:

    FL2V (Sage/Spectrum) — standard first/last-frame image-to-video and text-to-video, with SageAttention and the Spectrum speed node. Ref2va (Sage/Spectrum) — reference-image-to-video; holds identity and composition better for straight image-to-video. LightX2V Turbo — 4–8 step generation using the lightx2v turbo LoRA. Larryvrh Turbo — 4–8 step generation using Larryvrh's turbo LoRA; sharper at fewer steps in testing. Lip Sync (ref2va) — drive a portrait to sing or speak from reference audio, with closely matched lip movement. Lip Sync (LightX2V) — the turbo variant of the lip-sync workflow.

    Model files (downloaded automatically into the right folders): qwen3vl_32b_minimax_h3 text encoder (15.7 GB), video VAE fp16 (5.21 GB), audio VAE fp32 (605 MB), fl2va diffusion (21 GB), ref2va diffusion (21 GB), Larryvrh turbo LoRA (780 MB), lightx2v turbo LoRA (1.96 GB).

    The shared models — text encoder and both VAEs, ~21.5 GB — come with every menu option, since every workflow needs them.

  • System requirements

    OS: Windows 10 or 11, 64-bit.

    GPU: NVIDIA RTX 30XX / 40XX / 50XX. 24 GB VRAM is tested and comfortable; less can work using the memory-management nodes built into the workflows.

    Disk: at least 90 GB free if you choose "Everything" (models + ComfyUI + working space). Considerably less for the leaner menu options — Ref2va only needs roughly half that.

    Handled for you: Git, 7-Zip and the correct CUDA 12.8 PyTorch build are installed automatically. No manual environment setup.

  • Getting started

    1. Put the install .bat in its own dedicated folder and double-click it. Use a simple local path like C:\MiniMaxH3\ — avoid Desktop, Documents, and any OneDrive or Dropbox-synced folder, which can interfere with the install.

    2. When the download menu appears, press A, B, C, D or S for your choice. The window waits for your input and will not time out.

    3. Load any of the six workflows into ComfyUI and confirm the models are selected in the loaders.

    4. For the turbo workflows, keep steps at 4–8 (6 is a good balance) and LoRA strength at 1.0.

    5. For lip sync, use a clear front-facing portrait and clean reference audio, and match the output aspect ratio and duration to your source.

  • Troubleshooting

    Install fails or behaves oddly: check the install path first. A synced folder (OneDrive, Dropbox) or a deep path with spaces is the most common cause. Move to C:\MiniMaxH3\ and re-run.

    A large download dropped: just re-run — downloads are resumable and pick up where they stopped rather than starting over.

    Out of memory on generation: the workflows ship with memory-management nodes. Enable them, and drop resolution or duration before anything else. The turbo workflows at 4–8 steps are considerably lighter than the standard ones.

    A model isn't showing in a loader: confirm the menu option you chose actually included it — Turbo only (C) omits the ref2va diffusion model, and Ref2va only (D) omits fl2va. Re-run the installer and pick a fuller option if needed.

    Turbo output looks soft: raise steps toward 8 and confirm LoRA strength is at 1.0.

  • Licence & credit

    MiniMax H3 is by MiniMax. The turbo LoRAs are by the lightx2v team and Larryvrh. Custom nodes are by their respective authors — Kijai, rgthree, xmarre, Larryvrh, Kosinkadink and others.

    This product is the installer and workflow packaging — it automates the setup of their work into a one-click process. Model weights are downloaded from the original Hugging Face repositories at install time and remain under their own licences. Check the model's Hugging Face page for terms before commercial use.

$4.00Price
Quantity
    bottom of page