H3 Speed Boosters (optional packs)
The speed knobs that are safe to leave off
- model
- MODEL
H3 Speed Boosters is a panel of optional acceleration switches - Spectrum, TeaCache, block cache, and ComfyUI's built-in EasyCache - each measured, each behind a toggle, and each honestly described with its cost. The pack's own framing is the best summary: every booster changes the image slightly, like changing the seed, rather than degrading it - test on your own content before a long render.
The reason to care: a chained H3 render is expensive, and these are the legitimate levers. But the panel is also a masterclass in this pack's honesty - it tells you which boosters distort people, which do nothing at your step count, and which combination is a waste. You don't have to dig through release notes to know the block cache is inert at 14 steps; the tooltip says so.
How it works
You feed it the loaded model, flip the switches you want, and it applies the corresponding patches to a model clone. Boosters whose pack isn't installed print an install link and do nothing - the graph still runs, just at normal speed. The toggles:
spectrum- about 29% quicker, but can distort people in side-by-side comparison (rooms stay fine). NeedsComfyUI-Spectrum-MiniMax-H3.teacache+teacache_strength- about 14% quicker at the 0.15 default, 21% at 0.30; can distort people like Spectrum. Higher strength = faster and further from baseline. NeedsComfyUI-MiniMaxH3-TeaCache.blockcache+blockcache_threshold- skips model blocks whose input barely changed. The honest caveat: at the shipped 14 steps it never fires (measured 0 hits on two cards), because 14 steps are too far apart for its 0.12 threshold. Only worth it at 30+ steps. Ships OFF. Needscomfyui-minimax-h3-blockcache-T8.easycache- about 21% quicker and built into ComfyUI; nothing to install. Same advice: compare one scene on and off.
The output is the patched MODEL. A measured note in the node description kills the classic trap: stacking Spectrum + TeaCache is no faster than Spectrum alone - pick one.
Installing it
The node is part of the ComfyUI-H3-Multishot pack:
cd ComfyUI/custom_nodes
git clone https://github.com/jlucasmcrell/ComfyUI-H3-Multishot
Or search H3 Multishot in ComfyUI Manager. The booster packs themselves are separate optional installs - the node prints links for whichever you toggle without having. Requires ComfyUI v0.30.0+.
Gotchas
The panel's whole design philosophy is "safe to leave off," so the gotcha is the opposite of most nodes: don't stack. Spectrum plus TeaCache buys nothing beyond Spectrum alone, and you've paid the drift of two boosters for the speed of one. And don't turn on block cache at the shipped 14-step settings expecting a win - it's measured dead there; it only starts firing at 30+ steps. Finally, "distorts people" is not hyperbole - these are caching tricks that approximate attention and skip work, and human faces are where the approximation shows. Do the side-by-side the tooltips keep recommending before you commit a long chain.
Inputs (7)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| spectrum | BOOLEAN | false | About 29% quicker, but can distort people in side-by-side comparison (rooms stay fine). Try it on your scene before a long render. Needs ComfyUI-Spectrum-MiniMax-H3 installed. |
| teacache | BOOLEAN | false | About 14% quicker at default strength, 21% at 0.30 - can distort people, like Spectrum. Test on your scene first. Needs ComfyUI-MiniMaxH3-TeaCache installed. |
| teacache_strength | FLOAT | 0.150–0.5 | Higher = faster and the image drifts further from what you'd get without it. 0.15 is the measured default. |
| blockcache | BOOLEAN | false | Skips model blocks whose input barely changed. At the shipped 14 steps it never fires - measured 0 hits on every run on two cards - because 14 steps are too far apart for its 0.12 threshold. Only worth turning on at 30+ steps. Ships OFF. Needs comfyui-minimax-h3-blockcache-T8 installed (without it: a printed link, normal speed). |
| easycache | BOOLEAN | false | About 21% quicker and built into ComfyUI - nothing to install. Same advice as the others: compare one scene with it on and off. |
| blockcache_threshold | FLOAT | 0.120–1 | How different two consecutive steps may be before the block cache stops reusing work. 0.12 (the pack's default) never fires at 14 steps; higher values fire more often and drift the image more. Only meaningful with blockcache ON. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | — |