MiniMax H3 Easy Media Loader

MiniMax H3 is one of the first local video models that wants a lot of references at once - up to 9 images, 3 videos, and 3 standalone audio clips in a single generation. Wire those as nine separate Load Image nodes and your graph turns into spaghetti before you've even queued anything. That's the problem MiniMax H3 Easy Media Loader exists to solve: one node that holds your whole reference library - images, videos, audio - in an ordered list you can preview, reorder, replace, and remove from a single panel.

What it does

It's the input half of the ComfyUI-MiniMaxH3-Easy suite. Everything about MiniMax H3 is multimodal - text, image, video, and audio all share one context - so this suite built one visual manager for all of it. You drop files in, they show up as cards (video cards even display their duration), and the node hands the whole pile downstream as a single media_bundle.

That bundle feeds two places: the main MiniMax H3 Easy generation node, or MiniMax H3 Easy Context Segments when you're building a multi-shot video. Inside those, you reference specific items by typing @ in the prompt - @Image1, @Video1, @Audio1. Images, videos, and audio each number independently, so you can have all three at once without collisions.

How it works

Under the hood it's almost aggressively simple. The node has a single input, media_state (a string that holds your library's state), and a single output, media_bundle (MINIMAX_H3_MEDIA_BUNDLE). When you run, it loads each item from ComfyUI's input directory: images through the standard image loader, audio and video through the pack's own decoders.

The clever bit is caching. Media Loader computes a cache signature from each file's mtime and size, so the decoded frames of a reference video are reused through ComfyUI's node cache on later runs instead of being decoded from scratch every time. Once you've prepped a reference video once, subsequent generations skip most of the loading cost. Change or replace the file and the signature changes, invalidating the cache automatically.

One honest caveat: this cache is only in the Media Loader. The suite's other two input methods - direct media-port links and the Media Bridge - don't cache decoded video. And the cache saves prep time, not sampling time. The model still does its full denoising pass either way.

What to watch

The main gotcha is RAM. Decoded frames of a long or high-resolution reference video are held in system memory, so a big library can eat a surprising amount of it. Keep reference videos reasonably sized if you're on a tight machine. Also remember the limits: the regular generation path caps at 9 images, 3 videos, and 3 audio (15 media items total). The Context Segments shared library is more generous - 27 images, 9 videos, 9 audio - but each individual segment still obeys the per-call limits.

Installing it

This node ships in the ComfyUI-MiniMaxH3-Easy pack. First, update ComfyUI to a version that includes the official MiniMax H3 nodes - the pack builds on those rather than replacing them. Then either search ComfyUI-MiniMaxH3-Easy in ComfyUI Manager (pick the Nightly version - the README is explicit that Nightly is the current release and published tags may lag) or:

cd ComfyUI/custom_nodes
git clone https://github.com/nkxx188/ComfyUI-MiniMaxH3-Easy.git

Restart ComfyUI. The only hard dependencies are requests and psutil (plus FFmpeg if you use the segment-decode nodes). Models go in the usual diffusion_models/, text_encoders/, and vae/ folders, and the Easy Loader node picks them up directly.