MiniMax H3 PackageData

H3's signature move is that it understands text, image, video and audio as one context - and it's at its best when you hand it real material to work from: a character photo, a reference clip, a voice. This node is the basket you put all of that in. It gathers images, videos, and audio from two places - whatever you wire in, plus the external files you load through its front-end panel - and combines them into a single PACKAGE_DATA object that Conditioning or a Refiner can consume. One wire out instead of a spaghetti of media sockets.

How it works

Behind the scenes the node keeps a UI-side store (registered via /minimax-h3/package-data) where you can load external images, videos and audio by file. On the canvas it also accepts three optional typed inputs - images (IMAGE), videos (wildcard *, so video references from other nodes fit), and audios (AUDIO). When it runs, it merges the UI-loaded references and the wired inputs into one PACKAGE_DATA structure, each item getting a label (<Picture 1>, <Video 1>, <Audio 1>...). Those labels are exactly what prompt authoring then references - that's how the Storyboard's "a woman in the appearance of <Picture 1>" works.

There are hard caps in the source: 9 images, 3 videos, 3 audio references, with video and audio durations validated between 2 and 15 seconds (15 seconds total). Exceed them and the node refuses rather than silently dropping things.

The inputs and outputs

• images - optional IMAGE tensor (e.g. from a Load Image node).
• videos - optional, wildcard type; video-frame batches from whatever loader you use.
• audios - optional AUDIO.

One output: PackageData, type PACKAGE_DATA. It feeds either the MiniMax H3 Conditioning node's package input (for direct reference-based generation - the Ref2VA path in the README) or the Refiner nodes' package input (so the refiner can see and describe your references).

Installing it

Pack-wide routine - ComfyUI Manager search "MiniMax H3", or:

cd ComfyUI/custom_nodes
git clone https://github.com/xiaolibai-sys/ComfyUI-MiniMaxH3
pip install -r requirements.txt

then restart. No model files needed by the node itself.

Common issues

• Reference media ignored. Check the mutual-exclusion rule: PackageData and FL Constraint can't both feed the same Conditioning node. If both are wired, Conditioning refuses.
• Duration validation errors. H3 references need 2–15 seconds. A 1-second video clip gets rejected with a clear error - trim it up.
• Wired inputs and panel files both present. That's supported - they merge. But if you forgot which you added where, the output can surprise you. The panel shows a summary of what's loaded.
• Refiner doesn't see your images. If the Refiner node's package socket is empty, wire this node's output into it - and make sure the refiner's supports_image flag is on (that's the OpenAI-Compatible Refiner's job; the official Context IR refiner handles media itself).