Muse MiniMax H3 Unified Loader (POC)

MiniMax H3 is not one file you drop in. It's a ~33B omni-modal diffusion model plus a CLIP-style text encoder (possibly with an mmproj projector), a separate video VAE, a separate audio VAE, and - if you want it to fit on a normal card - a small pile of attention and memory patches. Wire all of that by hand and you're juggling five loaders and hoping the combo you picked matches your VRAM. This node, from Muse Collective (the group behind the Muse MiniMax H3 Director workflow), collapses the whole stack into one node with a profile dropdown. It's branded POC and a few weeks old, with essentially zero community footprint yet - but the idea is genuinely good: tell it your tier once, and it builds the stack.

How it picks for you

The profile dropdown offers Automatic, Low VRAM, Balanced, or Maximum Quality. On Automatic it reads your GPU's total VRAM at runtime and applies hard cutoffs: up to 12.5 GB gets Low VRAM, up to 20.5 GB gets Balanced, anything above gets Maximum Quality. Each profile has its own model + CLIP + mmproj slots, so you configure three quality tiers once and switch between them (or swap GPUs) without rewiring anything.

Two details are worth knowing. Model loading routes through ComfyUI-H3-Multishot's H3ModelLoaderAny, which measures real VRAM usage per render shape and tightens its estimate over repeat runs - not a fixed guess. And if the recommended profile has no model+CLIP pair configured (all slots on (disabled)), it falls back to the nearest configured tier and says so in the status text. If no profile is complete, it raises "Configure at least one complete MODEL/CLIP profile pair."

The inputs that matter

The dropdowns populate from your ComfyUI folders - models from diffusion_models, clips and mmproj files from text_encoders, VAEs and LoRAs from their usual places - so they'll read (disabled) until you've actually downloaded the files. A model+CLIP pair with mmproj on (auto) is the whole "must configure" story. Beyond that:

• low_vram_fl2va / balanced_fl2va / maximum_quality_fl2va - optional First/Last-frame model, one per profile, for workflows that need H3's hybrid-continuation branch. Leave disabled otherwise.
• lora_1..3 with strength and a target of Reference model, First/Last model, or Both models - the same LoRA can hit one or both model branches.
• attention_enabled - master bypass for every attention and feed-forward patch below it. Flip this off and all the SageAttention/low-VRAM/chunking settings below it are ignored, leaving you a clean MODEL to patch however you like.

The default patch set (SageAttention, low-VRAM attention with attention_head_chunks, and chunked feed-forward all on) is what makes a card actually render this thing - those defaults exist because H3 will OOM most consumer GPUs without them.

What comes out

model, clip, vae, audio_vae, and (if used) model_fl2va wire into the matching inputs on the Muse MiniMax H3 Director node - the intended destination. The model/model_fl2va sockets are plain MODEL outputs, so you can chain your own attention nodes in between. You also get three reporting outputs: selected_profile, detected_vram_gb, and a plain-text status listing what was actually loaded and patched this run.

Installing it

ComfyUI Manager (search Muse MiniMax H3 Unified Loader) or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/muse-collective-26/Muse-MiniMax-H3-Unified-Loader

Then restart ComfyUI. Now the part that will bite you: install ComfyUI-H3-Multishot and ComfyUI-KJNodes yourself. Manager's "install missing custom nodes" prompt won't catch them, because this node looks those classes up at runtime instead of placing them in the workflow graph - the only thing Manager's scanner reads. KJNodes is only needed while the attention patches are on (they are, by default); ComfyUI-H3-Multishot is always required. If you pick a .gguf model file you'll also want ComfyUI-GGUF.

Where people get burned

The classic failure is a fresh install running the node and getting RuntimeError: Required node 'H3ModelLoaderAny' is not installed or did not load. Newer versions of the loader name the exact package in the error, so read it before you panic - it's telling you to install ComfyUI-H3-Multishot. Other easy misses: only one half of a profile configured (you'll silently get a different tier - check status), or empty dropdowns because files aren't in the right folders.

One thing to know before you invest hours: the loader itself is MIT, but the H3 weights it feeds sit under the MiniMax H3 Community License, which is void in the US, EU, UK, and South Korea - outputs included. If you're in those regions, the legal local path doesn't exist yet, no matter how well the loader works.