Model Loader Pack

easy modelLoaderPack is the most boring node in this pack and it does exactly one useful thing: it takes a MODEL, a CLIP, a VAE, and an optional audio_vae, stuffs them into a single FAST_MODEL_LOADER bundle, and outputs one socket. The name isn't marketing - "pack" is the whole job. It exists because easy multitrackProject needs model, CLIP, video VAE and audio VAE to run H3, and dragging four fat cables across your graph every time is silly. Bundle once at the top, feed one wire.

How it works

There's no mechanism to speak of - and that's the point. The node validates nothing, loads nothing, and downloads nothing. It assembles a plain Python dict of {model, clip, vae, audio_vae?} and hands it out as a FAST_MODEL_LOADER. The "FastUse nodes" in the description are the same family of nodes inside this pack (the MiniMax H3 graph built by easy multitrackProject consumes that bundle). It's a data-flow convenience: a type-safe way to say "this whole cluster of model components travels together."

Where people get tripped up: audio_vae is optional, and H3 video generation genuinely needs it. H3 isn't a video model with audio bolted on - it generates stereo sound jointly with the picture, so the audio VAE is a real part of the bundle. If you leave it disconnected, the project node can't produce sound, or fails when it tries to build the joint audio/video latent.

The inputs and the output

Four inputs: model (the H3 diffusion model from your loader), clip (for text conditioning), vae (the video/image VAE), and audio_vae (optional, but treat it as required for video work). One output: model_loader, of type FAST_MODEL_LOADER. That's the wire that goes into easy multitrackProject's model_loader input (and optionally its model_loader_2nd input for a separate second-pass model).

A useful subtlety for the two-pass case: the project node lets you connect a second loader (model_loader_2nd) for pass two, but encoding and VAE components always come from the first-pass loader. So when you build your second bundle, you mostly care about its model - the CLIP and VAEs in it are ignored. Knowing that saves you a confused 20 minutes wondering why changing the second bundle's VAE did nothing.

Installation

It's part of ComfyUI-Easy-Media, so install the pack:

# FFmpeg installed first (the README insists)
cd ComfyUI/custom_nodes
git clone https://github.com/yolain/ComfyUI-Easy-Media

Restart ComfyUI, or use ComfyUI Manager (search "ComfyUI-Easy-Media"). No models or extra deps for this node itself - it's a dict packer. The real model load you'll be feeding it is the MiniMax H3 set, which is a separate, large download (~42.5GB for the full weights, and note the H3 community licence excludes the US, EU, UK and Korea from local use).

Troubleshooting

• "model_loader must contain a FAST_MODEL_LOADER dictionary" - you fed the project node something that isn't this node's output (or bypassed it). Connect model_loader from this node.
• Video renders with no audio / audio stage fails - audio_vae is empty. H3 needs it. Wire the H3 audio VAE.
• "currently supports only MiniMaxH3 models" downstream - the bundle is fine; the project node just won't run non-H3 models. Use the editor + a normal H3 graph instead.

It's glue. But glue is exactly what keeps a 30-node H3 project graph from turning into spaghetti, so it earns its place on the canvas.