MiniMax-H3-Multishot-Workflow

A ComfyUI node pack and two workflows that render a multi-shot MiniMax-H3 scene as one continuous take: no visible cut at the shot boundaries, no colour shift between shots, and continuous audio across the whole piece.

MiniMax-H3 natively generates blocks of roughly 10-15 seconds. This pack chains those blocks into arbitrarily long scenes and hands back a single master video with a single master audio track.

Released as v2.7.0 of the ComfyUI-H3-Multishot pack.

v2.7.0 - per-subject voices, flf_chain fixed, chain leveller, ComfyUI 0.34

• Per-subject voice refs (voice_ref_2 / voice_ref_3, both samplers): each character keeps their own voice across a chained scene - verified blind across a full chain, no cross-speaker bleed.
• flf_chain fix: the memory bank stood in flf chains and pinned a clip of shot 1 (which opens on boundary plate 0) into every later shot as an unbound reference - PLATE0 bled back in from shot 2 onward. The bank now stands down automatically in flf_chain; plates alone carry the continuity there.
• H3ChainNormalize: post-chain texture + colour leveller - the 2.6.0 long-take ratchet limit is addressed.
• refresh_pin splice alignment, x0 clamp dial (x0_clamp_window, dose cap 0.30), in-loop latent upscale + refresh_renoise + pin_noise_ramp + auto_chunk_ffn.
• ComfyUI 0.34 supported: native interior keyframe anchors detected; the pack's layout patch stands down automatically on 0.34, unchanged on older cores.
• Engine-aware writer: separate H3 ([English] ...) and LTX (quoted dialogue) system prompts for every mode via an appended engine widget.

v2.6.0 adds the extend take: one prompt, one continuous speech, as long as you want

Set take_seconds on MASTER CONTROLS (or open H3_Extend_Take, now the main workflow, shipped at 1280x736 with a 30-second take) and give the writer one premise. The panel sizes a window for your card and the number of windows that fills the time; the writer's new extend take join style writes ONE continuous speech and cuts it across the windows at sentence boundaries; the chain continues it under context_pin in H3's own voice — no TTS, no airlock, no per-shot dialogue budgeting. Verified before shipping: seven writer-driven renders including a 65-second, 7-window take, every join continued the speech, reviewed blind as one uninterrupted take. Also in 2.6.0: reference photographs now win over the writer's prose (the writer points at the photos instead of describing a face - same-seed measured), auto refs reach the sampler on their own (a second gate downstream used to discard them silently), three reserve-planning fixes from the 24 GB test lab and a writer fix for silent shots echoing the voice anchor. Full notes in the changelog below.

Known limit (2.6.0): the chain's texture ratchet is not fully solved for long takes - measured about +13% fine texture per join at 736x1280 with the anti-drift set on. Under ~4 windows (~30-40 s) it is slight; at 7 windows it is visible sharpening. Keep extend takes to ~4 windows for now; a pin-side fix is in progress for 2.6.1.

v2.5 is the memory release

Four new memory systems, all measured, two of them fully automatic:

• The driver-headroom rule (automatic). High-resolution renders used to take anywhere from 27 minutes to 3 hours for identical work — a lottery caused by the Windows driver demoting GPU memory when the card fills past roughly 95%. The pack now detects that zone and deliberately streams a few GB of weights instead of riding the ceiling. The lottery render became 15 minutes, every time. Nothing to configure.
• low_ram_master. Long chains used to hold every finished shot in system RAM until the final join — tens of GB at the very last step. Switch it on and shots stream to lossless disk staging as they finish; peak RAM becomes about two shots regardless of chain length, with verified-identical output (42.8 dB against the RAM path — codec noise).
• Remote text encoder. The text encoder runs for seconds per shot and holds 15+ GB the rest of the time. Point the new node at any second PC running ComfyUI with this pack and that memory leaves your render card — identical results, verified across machines, with a local cache so repeated text never touches the network. One flag on the switches panel turns it on; it ships off.
• H3 TAE Decode. 2-second full-resolution draft previews from a 9 MB tiny decoder, versus about a minute per shot through the real VAE. For seed hunts and batch triage, never finals.

Plus a Speed Boosters panel (Spectrum, TeaCache, block cache, EasyCache — each measured and eye-tested, with honest notes about which distort people), and new defaults tuned for 16-24 GB cards. Details in the changelog below.

What this repository contains

• ComfyUI-H3-Multishot/ - the ComfyUI custom-node pack (samplers, loaders, studio controls, LoRA stack, GGUF architecture patch).
• workflows/ - three ready-to-load ComfyUI graphs:
  • H3_Seamless_Chain_v2.json - the full workflow: master controls, LLM prompt writer, speed boosters, remote encoder lane, anchors.
  • H3_Seamless_Chain_CORE.json - the same chain with zero third-party dependencies.
  • H3_Keyframes.json - a single clip with anchors at chosen frame positions.
• INSTALL.md, SETTINGS.md, PROMPTING.md - install steps, the full settings reference, and the boundary/prompt rules.

What this repository does not contain

No model weights. Nothing here is a checkpoint, a text encoder, a VAE or a LoRA. Download the weights separately:

GGUF sizing guid