Asset → Slots 🐑

Here's the boring problem this node solves, and it's the kind that quietly ruins a render: your prompt plan says "Picture 2 walks into frame," and somewhere in the graph you have to hand-wire the right reference image to the right socket. Four characters, two scenes, and one mis-click later, the plan is describing one person's wardrobe on another person's face.

Asset → Slots 🐑 makes that mis-click structurally impossible. It takes the roster of asset bundles that the H3 loop plan was actually built from and splits it back into numbered outputs in that exact order - picture_1 is the image the plan calls <Picture 1>, because both came from the same slot. No matching, no counting, no hoping.

How it splits

There's exactly one input: assets (MIE_ASSET_ROSTER), fed from MiniMax H3 Loop Plan Generator V2's assets output. That's the only way in, by design - if you're not using the generator, put an Asset → Assets 🐑 collector in front of it and feed this node from there.

Then it fans out. picture_1 through picture_9 are IMAGE outputs, one per rostered asset, in slot order - characters first, then scenes, the same rule the generator uses when it numbers <Picture N>. audio_1 through audio_3 are AUDIO outputs carrying the characters' voice samples in character order; scenes never have audio, and the count stops at three because that's the cap on standalone reference audios in the H3 chain. Two integer outputs, count and audio_count, tell you how many of each are actually live.

Slots past the end of the roster emit None, and that's fine - wire any subset you want and leave the rest dangling. (The H3 ref2v node skips a None without consuming a reference number, so an empty slot doesn't shift your refs.)

One design detail is worth knowing because it looks like sloppiness otherwise: the output list is static. All nine pictures and three audios are always there, even when you have two assets. The pack tried trimming unused slots and hit the classic ComfyUI trap - removing outputs shifts link indices, and existing wires get silently retyped onto the wrong output, live-caught in this case as a ref_audio socket receiving IMAGE. So the extra sockets stay visible forever. Ignore them.

Where the outputs go

picture_1..N go to the reference-image inputs of your video conditioning node (ref_image_* on the H3 chain). audio_1..3 go to ref_audio_*. That's it - this is a wiring adapter, not a generator.

The reason to care is the same reason character consistency work keeps circling back to reference-driven pipelines at all: the model only holds an identity because it's looking at the reference while it generates, not because a sentence described it well. Give it the wrong picture and you get a confident, well-lit stranger. Give it the same picture the plan was written about, and the prompt and the conditioning agree by construction.

Install

ComfyUI Manager, search ComfyUI_MieNodes, or:

cd ComfyUI/custom_nodes
git clone https://github.com/MieMieeeee/ComfyUI-MieNodes
pip install -r ComfyUI-MieNodes/requirements.txt

No model downloads for this node. You do need images in the asset library for the roster to be non-empty, which means the Save Asset nodes (or some hand-placed folders) upstream, and the H3 video model downstream.

Things that go wrong

assets must be the generator's MIE_ASSET_ROSTER output, got X means you wired an IMAGE or a plain string into the single input - this socket only accepts the roster type, and it says so instead of guessing.

duplicate asset in the roster and the nine-picture cap both fire at execution, not on the canvas. Same asset in two slots is always a copy-paste: one asset, one slot.

And the quiet one: if the count of pictures the plan references is larger than the count you actually wired, nothing errors - you just get scenes referencing <Picture 5> when only three exist. Check count against your plan's scene count before you queue a long render; it's cheaper than discovering it at clip four.