One node. Write a shot plan, drop in your reference stills, queue once.

Hand Tie Clips renders a multi-hop MiniMax H3 Reference-to-Video chain with

audio. Each hop is one continuous clip, and consecutive hops are joined by

pinning the previous hop's final frames and audio tail into the next hop's

conditioning — so a 30-second scene comes out of one queue instead of six

manual runs stitched together in an editor.

# Read this before you download #

This workflow ships the turbo stack — the LoRA loader, the AdaLN fix, sparse

attention and the low-VRAM path. That stack is why it runs at 4–8 steps instead

of 14, and it means the graph needs three node packs, not one. Without them

it opens with red missing-node boxes.

### Node packs ###

| [ComfyUI-Hand-Tie-Clips](https://github.com/dntpi/ComfyUI-Hand-Tie-Clips) | this workflow's node — the chain, the seam report, the chain preview |

| [ComfyUI-PlagueKind-Nodes](https://github.com/PlagueKind/ComfyUI-PlagueKind-Nodes) | the turbo LoRA loader, AdaLN LoRA fix, sparse attention |

| [ComfyUI-KJNodes](https://github.com/kijai/ComfyUI-KJNodes) | low-VRAM attention, latent preview override |

| [ComfyUI-H3-Motion-Context](https://github.com/NikoDemon80/ComfyUI-H3-Motion-Context) | optional but recommended. Hops after the first are guided by the previous hop's sampler latent through it. Without it the node falls back to stock MiniMaxH3AddGuide — still works, still chains, different join. |

Install Hand Tie Clips itself with:

cd ComfyUI/custom_nodes

git clone https://github.com/dntpi/ComfyUI-Hand-Tie-Clips.git

Then restart ComfyUI and hard-refresh the browser (Ctrl+Shift+R). A stale

browser cache is the single most common reason the node mounts with no UI.

There are no Python dependencies to install — everything it imports already

ships with ComfyUI. A zip is attached for anyone who would rather not clone,

but git clone is the better path: updating later is just git pull.

### Models ###

These are the filenames my loaders point at. **If your H3 weights are named

differently, just repoint the loaders** — nothing here is hardcoded.

UNET minimax_h3_hybrid_fl2va_ref2va_b30-49-int8.safetensors

CLIP qwen3vl_32b_minimax_h3_int8_convrot.safetensors (type: minimax)

VAE minimax_h3_video_vae_int8_convrot.safetensors

VAE minimax_h3_audio_vae_fp32.safetensors

LoRA minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors

TAE taeh3.safetensors (preview only)

You need a ref2va or hybrid ref2va checkpoint — plain fl2va has no

reference rows and will not work.

## What it's for##

- Scenes longer than a single H3 generation

- Keeping a character's face, wardrobe and voice consistent across hops

- Re-rolling one hop without re-rendering the rest

## What's in it ##

| shot_plan | your script as JSON — one shot per hop, so the shot count is the hop count |

| directives | join / camera / framing / pace / tail per shot, compiled to vetted prose |

| ref_plan | stable @tags for reference stills, grouped into subjects |

| cache_hops | lossless per-hop cache — edit hop 5 of 8 and only 5–8 re-render |

| seam report | measures the brightness step at every join, in /255 |

| tone compensation | corrects cumulative brightness drift across a long chain |

## Getting started

Once the packs above are installed, open the attached

HandTieClips_Starter.json. It ships with no reference images on purpose,

so it runs before you've supplied any pictures. The canvas carries its own

documentation as note cards.

prompt_pack/ in the pack folder turns any chat model into a plan writer: paste

SYSTEM_PROMPT.md into LM Studio's system prompt box, describe your scene in

plain language, and paste the two JSON blocks it returns straight into the node.

## Honest limits

- Built and tuned for the turbo regime — 4 to 8 steps, not 14.

- A 5-second hop drops the airlock on a continuous join. Validate seams at 8s or 15s.

- Written for MiniMax H3 specifically. It is not a general video-chaining tool.

Full docs, prompting guide and engineering log:

https://github.com/dntpi/ComfyUI-Hand-Tie-Clips

## Writing the plan (optional, but this is the good part)##

shot_plan and ref_plan are ordinary JSON, so you can type them yourself and

the node will tell you exactly what it doesn't like. But the pack ships

prompt_pack/ to turn any decent chat model into a plan writer:

1. Open LM Studio (or any local/hosted chat model), paste the contents of

prompt_pack/SYSTEM_PROMPT.md into the system prompt box.

2. Describe your scene in plain language — *"three 10-second clips, a cook in a

kitchen, she finishes plating and walks out into the hallway."*

3. Paste the two JSON blocks it returns straight into the node.

It's worth doing even if you're comfortable with the JSON. The format has two

rules that are easy to get wrong by hand and annoying to debug: a reference

@tag has to match its register entry character for character, and beats must

describe what is happening rather than what stops happening — H3 renders the

negation. The prompt doc encodes both.

Tested against Qwen and Gemma-class local models; nothing exotic required.