A 30-Second Multi-Shot Chain for ComfyUI: The Ratchet Number the Model Card Buries

What happened

LaDruid published v2.7.0 of the ComfyUI-H3-Multishot node pack to Hugging Face on 19 September 2026. It is a set of ComfyUI nodes and three workflow JSON files that chain MiniMax-H3's 10-to-15-second video blocks into one continuous take with a single master audio track.

Context

MiniMax-H3 generates short blocks — 10 to 15 seconds — natively. Sticking them together before this pack meant external editing, per-shot audio matching, and visible seams at every join. The pack has iterated through v2.5 (memory management: driver-headroom detection, low_ram_master disk staging, a 9 MB TAE preview decoder, remote text-encoder offload), v2.6.0 (the H3_Extend_Take workflow at 1280×736 with a 30-second default), and now v2.7.0, which layers per-subject voice references, a post-chain texture and colour leveller, and a memory-bank fix on top of that foundation.

How it works

The pack plugs into an existing ComfyUI installation — v2.7.0 supports ComfyUI 0.34 and auto-detects native interior keyframe anchors, standing down its own layout patch on 0.34 while leaving older cores unchanged. It bundles no MiniMax-H3 weights. You load one of three workflow JSON files: the full node tree (H3_Seamless_Chain_v2.json), the zero-third-party-dependency CORE variant (H3_Seamless_Chain_CORE.json), or H3_Keyframes.json. The samplers chain native H3 blocks while a shared memory bank carries texture, colour, and audio forward across shots.

v2.7.0 adds per-subject voice references (voice_ref_2, voice_ref_3) in both samplers so each character retains their own voice across the entire chain, verified blind with no cross-speaker bleed. H3ChainNormalize post-processes the stitched output to level texture and colour drift. An x0_clamp dial (dosed to a 0.30 cap) controls how aggressively the chain resists drift. The engine-aware writer selects between H3's tag-wrapped prompt format and LTX's quoted-dialogue format via an appended widget. Defaults are tuned for 16-to-24 GB GPU cards.

Our read

The number that matters is not the 65-second, 7-window verification render the model card leads with. It is the +13% texture ratchet per join, measured at 736×1280 with the anti-drift set enabled. Under four windows — roughly 30 to 40 seconds — the ratchet is slight. At seven windows it produces visible sharpening. The documentation recommends capping extend takes at four windows and flags a pin-side fix for v2.6.1 that had not shipped as of 19 September. The practical ceiling for deliverable-quality output is about 30 seconds of clean continuous footage, not 65. The demo proves the chain holds; the ratchet number says where it starts to look synthetic.

The model card also buries a structural fact: this is glue code. No MiniMax-H3 weights are included in the repository, the sources do not state where to obtain them or under what licence, and the base model is a hard prerequisite. Apache-2.0 covers the node pack and workflow files, but a studio evaluating this for a client deliverable needs to confirm the base model's own terms separately.

For a small studio with a 16-to-24 GB card, the v2.5 memory stack is where the real operational value sits. The driver-headroom rule detected the Windows GPU memory zone above roughly 95% utilisation and streamed weights to avoid driver demotion, reducing a lottery render that ranged from 27 minutes to 3 hours down to a consistent 15. low_ram_master streams finished shots to lossless disk staging so peak RAM stays at about two shots regardless of chain length, verified identical at 42.8 dB against the RAM path. The 9 MB TAE decoder produces a 2-second full-resolution draft versus roughly a minute per shot through the full VAE, which changes how you iterate on prompts before committing GPU time to a final render.

What this changes

If you already run ComfyUI and hold MiniMax-H3 weights, H3_Seamless_Chain_CORE.json is a drop-in: one file, zero third-party nodes, no dependency tree to resolve. Load it, wire your prompts, render. Cap chains at four windows until the v2.6.1 pin-side fix lands. Use the TAE decoder for prompt iteration — 2 seconds per draft instead of a minute — and switch to the full VAE only when the composition is locked. Enable low_ram_master on a 16 GB card if the chain exceeds two shots. The per-subject voice references eliminate the audio-stitching pass for multi-character scenes. If you do not already have a ComfyUI installation and the MiniMax-H3 base model, this pack alone changes nothing on Monday.

License

The ComfyUI-H3-Multishot node pack and its workflow files are released under Apache-2.0, as stated on the Hugging Face repository. Commercial use of the code and workflows in this repository is permitted without restriction. The licence does not extend to the MiniMax-H3 base model weights, which are not bundled here; the sources do not state that model's licence, so check the MiniMax-H3 model card before shipping anything built on it.

Key takeaways

• This is a ComfyUI node pack and three workflow files, not a standalone model. It requires an existing ComfyUI installation and MiniMax-H3 base weights that are not included in the repository.
• The practical ceiling for clean multi-shot output is roughly 30 to 40 seconds (four windows) due to a +13% texture ratchet per join; a fix is flagged for v2.6.1 but had not shipped at publication.
• Per-subject voice references (voice_ref_2, voice_ref_3), verified blind with no cross-speaker bleed, remove the audio-stitching step for multi-character scenes.
• Memory features are tuned for 16-to-24 GB GPUs: driver-headroom detection (27 min to 3 hr lottery → 15 min), low_ram_master disk staging (peak RAM ≈ 2 shots, 42.8 dB identical), and a 9 MB TAE decoder for 2-second draft renders.
• The CORE workflow variant carries zero third-party dependencies, making it a clean drop-in for a minimal ComfyUI setup.

Sources

1. LaDruid/MiniMax-H3-Multishot-Workflow — text-to-video on Hugging Face — tier 3