MiniMax-H3 Multishot — chained shots, one master, with audio — v1.0 — v2.6 Extend Take

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

Auto refs now reach the sampler on their own. The REFERENCE IMAGES lane had two switches in series - USE AUTO REFS, and a REFERENCE gate downstream that silently threw the auto refs away unless you also flipped it (the log said "3 ref(s)" and the character rendered from prose alone). USE AUTO REFS alone is enough now; the gate is the MANUAL REFS gate and only guards the two LoadImage slots. Proof line in the log: "N reference image(s) ride in every shot".

The writer no longer overrides your reference photographs. Same seed, photographs attached: the writer's identity sentence ("a woman in her thirties with dark hair tied back") rendered that person; the same prompt with the sentence replaced by "looks exactly as in the reference photographs" rendered the person in the photographs. The writer has a refs_attached input, wired from USE AUTO REFS in the shipped canvases; when on it points at the photos and describes nothing about face, hair, age or build.

Saved 2.5.x canvases keep working: the VRAM / SPEED panel dropped its five dead toggles but keeps all eight output slots in the original order (saved graphs link by slot number); the removed slots always emit false.

take_seconds = 0 (the default) is exactly the old behaviour; saved workflows are unaffected.

audio_pin_frames on the memory sampler: the audio reference window independent of the picture pin (96 = a 4-second audio memory; measured neutral so far, ships 0).

Reserve-planning fixes from the 24 GB test lab: no more phantom x1.6 on CORE-workflow reserves; first runs borrow measurements only from the same quant family (GGUF figures do not transfer to w4a8/int8); streamed masters embed the real graph again.

Writer fix: a silent shot whose boilerplate mentioned lip movement, with a voice anchor in the chain, could re-speak an earlier line. Fixed at the source, and the writer now warns by shot number.

2.5 — the memory release

Four new memory systems, all measured, two of them fully automatic. If renders have ever randomly crawled, if a long chain has ever killed your machine at the final join, or if the text encoder's 15 GB has ever been the difference between fitting and not — this release is about you.

New: the random-slowdown fix (automatic)

The single biggest fix this pack has shipped. High-resolution renders would randomly take anywhere from 27 minutes to 3 hours for identical work — same seed, same settings, a lottery. The cause: when the model weights plus working memory fill the card past roughly 95%, the Windows driver starts quietly demoting GPU memory, and whether your render crawled was luck.

The pack now detects that zone before sampling and deliberately streams a few GB of weights instead of riding the ceiling — streamed weights are nearly free on modern ComfyUI, the last few percent of VRAM are not. The lottery render became 15 minutes, every time . Nothing to configure; the console prints a "driver headroom" line whenever it saves you.

New: low_ram_master — long chains without the system-RAM cliff

Off (the default), every finished shot is held in system RAM until the final join — historically the point where long renders killed whole machines, after all the sampling was already paid for. Switch it ON and each shot streams to lossless disk staging as it finishes; the master is assembled from disk through the exact same levelling math. Verified at 42.8 dB against the in-RAM path, which is codec noise. Peak RAM stays near two shots no matter how long the chain is. Turn it on if you chain long or run under 32 GB of system RAM. The finished file's path comes out of the new master_path output either way.

New: remote text encoder — the 15 GB your render card never needed

The text encoder works for a few seconds per shot and occupies 15+ GB the rest of the time. If you have any second PC with ComfyUI: install this pack on it, point the new H3 Remote Text Encoder node at it, turn remote_encoder ON in the VRAM / SPEED SWITCHES panel, and prompts are encoded over there while your render card keeps the memory. Results are identical — verified across two different machines to float precision — and repeated scene text is answered from a local cache with no network call at all. It ships wired into the full workflow; the panel flag defaults OFF, so single-PC setups see zero change. A how-to note sits right next to the node.

New: H3 TAE Decode — 2-second draft previews

A 9 MB tiny decoder (Kijai's taeh3 , one small download into models/vae_approx/ ) turns latents into full-resolution draft frames in about 2 seconds, versus roughly a minute per shot through the real VAE. Drafts smear fine texture, but composition, framing and motion read clearly. Use it to audition seeds or triage a batch, then decode the keepers through the real VAE. Never for finals.

New: the Speed Boosters panel — every option, measured, with the honest verdicts

Spectrum, TeaCache, block cache and ComfyUI's own EasyCache, each behind a switch on one node. All measured on the same seed, and all eye-tested on the same finished videos:

block cache : no effect at 14 steps — 0 cache hits on every run on two cards (the earlier -11% did not reproduce). Ships OFF; only worth it at 30+ steps.

Spectrum : -29%, but visible distortion on people. Rooms and environments were fine.

TeaCache : -14% to -21% depending on strength, same story — people distort, environments hold.

EasyCache (built into ComfyUI, nothing to install): -21%, same family of people distortion.

Stacking Spectrum + TeaCache : no faster than Spectrum alone and severely damaged output. Never stack them.

If your scene has no people in it, the aggressive boosters are real money. If it has people, leave them all off. A booster whose pack is not installed prints an install link and passes the model through unchanged — it no longer breaks the graph.

Changed: defaults tuned for 16-24 GB cards

736x1280 at 192 frames and 14 steps, euler / beta57 , the curve-Q5_1 checkpoint (~14 GB), and the full anti-drift set on: context_pin , chain_gain_control=flatten , master_normalize=luma+contrast , pin_renorm , memory_frames=0 . The resolution floor is deliberate — the base model distorts faces below roughly 1 megapixel, so shrinking the canvas to save VRAM costs you faces first.

Changed: the writer now knows how long a shot really is

Chained shots spend about five seconds of every block on the replay, airlock and settle, so the speakable span is shorter than the clip. The prompt writer now receives that real span and sizes every line against it, and overruns or fully-silent scripts are flagged at generation time with the numbers. Measured on renders: lines over budget garble, lines under budget drag, matched budgets came back natural and fully intelligible.

Changed: writer craft rules learned from failed renders

Silent shots with visible people now state what mouths are doing — stops the model inventing mumbling.

Things revealed mid-shot are written as already present — stops pop-ins.

A shot that opens a mouth without dialogue must name the sound it makes.

New: clickable title blocks

Every workflow now carries its version number and live links to the GitHub repo, the HuggingFace page and this page, plus a plain-language pass over every on-canvas note. If you have ever wondered which version of a workflow you had open — that is over.

Every new control in this release is appended at the end of its node's input list, so workflows you saved against 2.2.x load unchanged.

Text encoder · Unknown