😺NKD MiniMax Guides
Every MiniMax H3 guide from one node, instead of a row of spaghetti
- positive
- vae
- audio_vae
- latent
- guides
- positive
- latent
- video vae
- audio vae
MiniMax H3 is an omni-modal model - text, image, video and audio are all one input context - which means a serious H3 shot has multiple guides: a reference still, a clip whose motion you want to copy, its soundtrack, maybe another frame sequence. The stock way to wire that is a row of "Add Guide for MiniMax H3" nodes, each dragging the same four cables in from far upstream. Move one and the canvas turns to spaghetti. 😺NKD MiniMax Guides is the same anchoring in one node, and the guide list grows as you fill it.
How it works
The node has a guide list that grows as you connect slots - each slot takes a still image, a frame sequence, a video (picture and its soundtrack), or bare audio. When you fill a slot, a position widget appears next to it for the frame it lands on; negative positions count from the end of the clip. Two slots sharing a position - a clip and its sound - become one guide, which is how picture and audio stay anchored together.
The plumbing win is that latent, video vae, and audio vae come straight back out of the node. The sampler and whatever comes next hang off this node instead of reaching back across the graph for cables that came in at the top. Both VAEs are required inputs rather than optional, which keeps them at the top of the node, above the growing guide list.
The inputs
positive- your CONDITIONING.vae/audio_vae- the video VAE and the audio VAE, for the slots that carry picture and sound.latent- the working latent.guides- the autogrowing slot list (aCOMFY_AUTOGROW_V3type).position_1throughposition_12- the frame anchor for each of the twelve possible slots, appearing as you fill them.
Outputs: positive, latent, video vae, and audio vae - everything the sampler needs, at the bottom of the node.
The anchoring itself is the core node's - same clip-length snapping, same audio cropping, same errors - so this isn't a parallel implementation that drifts, it's the stock behavior with the wiring collapsed.
Install and gotchas
Install via ComfyUI Manager (search ComfyUI-NKD-Basic-Tools) or:
cd ComfyUI/custom_nodes
git clone https://github.com/Nekodificador/ComfyUI-NKD-Basic-Tools
Restart after. Pack dependencies are onnxruntime and huggingface_hub; neither is needed here. This is an H3-specific node, so the standing H3 caveats apply: the weights are ~42.5 GB full precision, there's no verified consumer-VRAM floor yet, and the community licence excludes the US, EU, UK and South Korea - check your region before building a workflow around it. And if you're doing masked inpainting rather than multi-guide anchoring, this isn't the node you want - that's 😺NKD AV Latent, which handles the mask side.
Inputs (17)
| Name | Type | Default | Description |
|---|---|---|---|
| positive | CONDITIONING | — | |
| vae | VAE | Video VAE, for the slots that carry picture. | |
| audio_vae | VAE | Audio VAE, for the slots that carry sound. | |
| latent | LATENT | — | |
| guides | COMFY_AUTOGROW_V3 | — | |
| position_1opt | INT | 0-9999–9999 | Frame index this guide is anchored at. Negative counts from the end. Two slots sharing a position (a clip and its sound) become one guide. |
| position_2opt | INT | 0-9999–9999 | Frame index this guide is anchored at. Negative counts from the end. Two slots sharing a position (a clip and its sound) become one guide. |
| position_3opt | INT | 0-9999–9999 | Frame index this guide is anchored at. Negative counts from the end. Two slots sharing a position (a clip and its sound) become one guide. |
| position_4opt | INT | 0-9999–9999 | Frame index this guide is anchored at. Negative counts from the end. Two slots sharing a position (a clip and its sound) become one guide. |
| position_5opt | INT | 0-9999–9999 | Frame index this guide is anchored at. Negative counts from the end. Two slots sharing a position (a clip and its sound) become one guide. |
| position_6opt | INT | 0-9999–9999 | Frame index this guide is anchored at. Negative counts from the end. Two slots sharing a position (a clip and its sound) become one guide. |
| position_7opt | INT | 0-9999–9999 | Frame index this guide is anchored at. Negative counts from the end. Two slots sharing a position (a clip and its sound) become one guide. |
| position_8opt | INT | 0-9999–9999 | Frame index this guide is anchored at. Negative counts from the end. Two slots sharing a position (a clip and its sound) become one guide. |
| position_9opt | INT | 0-9999–9999 | Frame index this guide is anchored at. Negative counts from the end. Two slots sharing a position (a clip and its sound) become one guide. |
| position_10opt | INT | 0-9999–9999 | Frame index this guide is anchored at. Negative counts from the end. Two slots sharing a position (a clip and its sound) become one guide. |
| position_11opt | INT | 0-9999–9999 | Frame index this guide is anchored at. Negative counts from the end. Two slots sharing a position (a clip and its sound) become one guide. |
| position_12opt | INT | 0-9999–9999 | Frame index this guide is anchored at. Negative counts from the end. Two slots sharing a position (a clip and its sound) become one guide. |
Outputs (4)
| Name | Type | Description |
|---|---|---|
| positive | CONDITIONING | — |
| latent | LATENT | — |
| video vae | VAE | — |
| audio vae | VAE | — |