FeiHou Easy H3 Media

Open this node and you'll think ComfyUI failed to load it. No input sockets, one output, and a 3×3 grid of empty image slots with three video slots and three audio slots underneath. That grid is the node. The whole point of the FeiHou Easy H3 pack is that reference media loading happens inside the generation node instead of through a fan of Load Image, Load Video and Load Audio nodes, and this node is the half of that idea that got split off so the settings half can stay tidy.

What it is, and why it exists

H3 is a reference-driven model - the MiniMax H3 "reference to video" path takes up to nine images, video clips and audio clips as context, and its official node ships a media panel to hold them. The upstream pack this one adapts moved that panel inside its main node; the split edition pulls the gallery back out into its own node so a workflow can have a media node you can clone, mute, or park off to the side while the generation settings live in the node you're actually tweaking. Niche, but if you've ever fought a 900-pixel-tall node that won't let you see two rows of the graph at once, you know why it exists.

How it works, honestly

The Python side is four lines. FeiHouEasyH3Media declares no required or optional inputs, returns a single h3_media output of type FEIHOU_H3_MEDIA, and its function returns a marker dict. All the actual work happens in the pack's frontend JavaScript: you click a slot to pick a file or drag files onto it, images and videos preview inline, audio shows a filename, and the chosen files are uploaded into ComfyUI's normal input folder. Those uploads persist with the workflow, so reopening it restores what you had.

The media then reach the generator through hidden transport fields. FeiHou Easy H3 Setup declares 45 hidden media_*, media_type_* and media_trim_* string inputs, exactly the ones the original all-in-one main node uses; the gallery fills them on the way to the queue. This is why the node's own output looks so thin - the h3_media wire exists to make the connection real and to force Setup to run, and the payload travels out of band. Same trick as reference mentions inside the prompt: you type @ in the prompt and get <Picture 1>, <Video 2>, <Audio 1> tokens that line up with the slots.

Two limits are worth internalising. Reference mode needs at least one image or one video - audio alone won't run. And videos keep their own soundtrack paired to the video; the three independent audio slots are extra references on top, not the video's audio.

Install and wire-up

cd ComfyUI/custom_nodes
git clone https://github.com/FX-FeiHou/ComfyUI-FeiHou-Easy-H3
# restart ComfyUI

Or install "ComfyUI-FeiHou-Easy-H3" from ComfyUI Manager and restart. Nothing to pip install. This node only works against the split edition of the pack: h3_media is a required input on FeiHou Easy H3 Setup, and it only knows what to do with a FEIHOU_H3_MEDIA link.

Wiring is one cable. h3_media → Setup's h3_media socket. Then everything else is Setup's business: its h3_context output is what you feed to the dual-sample or sample-enhancer nodes, and model / second_sampling_model are what you feed to the samplers.

If it misbehaves

Most reports here will boil down to three things. If the graph won't queue, check the h3_media link - an unconnected required input is a hard stop, and the node visually looks "empty" so it's easy to miss. If your selections vanish after a reload, the files were uploaded but the frontend state didn't survive - re-pick, and check the browser console; this pack is frontend-heavy enough that the JS is the load-bearing part. And if you select media and Setup behaves as though there's none, you're probably running the original all-in-one main node instead of Setup, which never reads the h3_media socket - that input only exists on the split node.

One last thing, since these are reference assets: H3's weights ship under MiniMax's community licence, which excludes the US, EU, UK and South Korea from the grant entirely, outputs included. That's a licence question, not a node bug, but it's the kind of thing worth knowing before you build a pipeline on it.