沐阳 H3 · 选择角色参考图（内部）

The Director works with a media inventory - one shared bag of images, videos and audio that every shot in the job can draw on, plus per-shot media. Inside a prompt you point at a specific item with @图片1, @视频2, @音频1, and the Media Agent rewrites those into H3's official <Picture N> / <Video N> / <Audio N> tags.

Most of that machinery is string work. This node is the exception: sometimes a downstream consumer needs an actual IMAGE tensor, not a label. H3DirectorMediaImage grabs one image out of the bag and hands you pixels.

What it does

Two inputs, and they're both required:

• media - type MINIMAX_H3_MEDIA. That's the Director's or Media Agent's inventory object.
• image_ordinal (1-9, default 1) - which image you want, counting images only.

The counting detail matters. The ordinal indexes the image items in the pack, skipping videos and audio, so image_ordinal: 2 is the second image in the inventory even if there are three audio files sitting ahead of it. Three videos in the bag don't shift the numbering. If you've built workflows where every media type shared one index, that's the trap to unlearn - the pack gives each media type its own index on purpose, both here and in the @ tags.

One output: 角色参考图, a single IMAGE. It returns the first frame only (frames[:1]), so it's a still no matter what the source item was. Don't feed it a video and hope.

If you ask for an item that doesn't exist, you get a specific error naming the ordinal and how many images the inventory actually holds, which is more helpful than the average ComfyUI traceback. The error is telling you to add the image to the shared media or to the Media Agent, or to change the ordinal.

Where it fits

It's the fetch half of the multi-view path. The Director's 五视图 (five-view / turnaround) flow needs to pick a character reference out of the shared bag and feed it into a reference-image consumer; this is the node that does the picking, and H3DirectorTurnaroundMedia is the node that puts the generated turnaround sheet into the bag in the first place. They're a matched pair.

Category 沐阳 H3/导演台/内部 - internal, meaning the Director wires it for you in the normal flow. It's worth understanding anyway the moment you cross-reference @图片N labels with the actual tensors, because the two numberings are the same numbering and mismatching them is how you end up with a reference image of the wrong character.

One thing this is not

It doesn't encode anything. No CLIP, no VAE, no conditioning. It hands you a decoded image tensor and stops. Whatever consumes that image is responsible for the encode - the same separation the pack uses everywhere else, and the reason these internal nodes are cheap enough to leave in the graph.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/civilcoco/ComfyUI-MiniMaxH3-Myang

Restart ComfyUI, hard-refresh the browser. Find it under 沐阳 H3/导演台/内部, or search "ComfyUI MiniMax H3 Myang" in ComfyUI Manager.

No Python dependencies - the pack declares an empty dependency list, since the media logic is its own code over torch. The node itself relies on the media-catalog types shipped inside the pack, so it only works with a MINIMAX_H3_MEDIA produced by this pack's Director or Media Agent. Hand it something else and you get a validation error rather than a mystery.

Practical notes

If you're building a character-consistency chain for a long H3 job, the useful pattern is one strong face close-up per character in the shared media, referenced by ordinal everywhere it's needed, rather than re-uploading the same image onto every shot card. Shots get their own media or can append the shared inventory - the shared bag is the thing that keeps identity references consistent through a chain. The node plumbing layer is worth a read if you want to understand why a pack would route images through a typed inventory object instead of nine separate IMAGE sockets; and identity preservation covers why you want a reference at all.