🧩 H3 Loader - LD

Why you need four loaders when a normal video model needs two

MiniMax H3 is an omni-modal model: it treats video and audio as one generation context, and that means it doesn't ship a single VAE like an image model would. It has two - one for the picture, one for the sound - plus its own diffusion model and CLIP text encoder. So a bare-bones H3 graph starts with four loader nodes doing jobs you don't normally think about separately, and it's exactly the kind of thing that makes a clean tutorial graph look like a plate of spaghetti.

H3 Loader - LD (🧩 H3 Loader - LD, under LD / PromptMaster, in Lora-Daddy's PromptMasterLD pack) collapses that into one themed panel: the input block. It loads the UNET, the CLIP, the video VAE and the audio VAE in one node and hands you all four outputs from a single place.

What each slot actually is

• UNET - the H3 diffusion model itself. In ComfyUI you'll usually load it from a folder of H3 files (the same weight files the model card ships).
• CLIP - the text encoder that turns your prompt into conditioning.
• video VAE - H3's picture autoencoder.
• audio VAE - H3's sound autoencoder, the thing most tutorials forget and then wonder why their clips come back silent or their latent decode crashes.

The dials are straightforward: load_model on/off, unet_name, weight_dtype (float16 vs bf16 depending on your card), clip_name, clip_type, clip_device, and the two VAE dropdowns vae_video and vae_audio. H3's weights are big - expect to pick the quant that fits your VRAM and to run the audio VAE on CPU if you're tight.

The override trick that makes it an "input block"

The optional inputs are where this node earns the "block" name. Wire a MODEL into model_in and the loader skips loading one from disk, using yours instead - same idea for clip_in, vae_video_in and vae_audio_in. That turns H3 Loader from a fixed load point into the header of a rail you can drive from elsewhere: a LoRA'd model, a different CLIP, a VAE you've already got in memory for a second graph. If you never wire them, the node behaves exactly like four ordinary loaders.

Outputs are the obvious four: model, clip, vae_video, vae_audio. Those plug straight into any H3 sampler - H3 Multishot, H3 Chain, or the pack's own Master and Music Video nodes all take the same quartet.

How to install

This is a PromptMasterLD node, so:

cd ComfyUI/custom_nodes
git clone https://github.com/Lora-Daddy/PromptMasterLD
cd PromptMasterLD
pip install -r requirements.txt

Restart ComfyUI (or install via ComfyUI Manager by searching "PromptMasterLD"). The pack adds opencv-python as its one hard dependency; ComfyUI already ships the torchaudio and PyAV layers H3's dual-VAE decode needs.

The gotcha that gets everyone

The two-VAE thing is where H3 newcomers fall in. If you only wire the video VAE because you're used to image models, decoding an H3 latent that includes audio blows up or comes back silent. The whole reason this node exists is to make the four-piece loading a single visible habit. And a model wired into model_in takes precedence over unet_name - leave unet_name set anyway, so the node still works if you unwire the override later.