Multishot Advance Speed Boosters (optional packs)

H3 renders are slow, and by the time you're chaining multiple shots they're really slow. The usual fix is to bolt on community accelerator packs - TeaCache, Spectrum, block-cache - but each one means installing another node pack and cluttering your graph with nodes you have to wire in the right order. This node is the lazy person's version: it applies those accelerators to your model behind plain on/off switches, so you get the speed without the wiring.

Read the node's own description first, because it sets expectations honestly: every booster changes the image slightly - the author compares it to changing the seed - rather than degrading it. Test on your own scene before committing to a long render.

The switches

• easycache - about 21% quicker and built into ComfyUI, so there's literally nothing to install. Same "compare one scene with it on and off" advice as the rest. This is the one to try first.
• spectrum - about 29% quicker, the biggest single win. Needs ComfyUI-Spectrum-MiniMax-H3. Caveat: it can distort people in side-by-side comparison (rooms stay fine), so don't trust it on faces without testing.
• teacache - about 14% quicker at default strength, up to 21% at teacache_strength 0.30. Needs ComfyUI-MiniMaxH3-TeaCache. Higher strength = faster but drifts further from the unboosted image; 0.15 is the measured default.
• blockcache - the honest trap. It skips model blocks whose input barely changed, and at the shipped 14 steps it never fires - the author measured zero hits on every run across two cards, because 14 steps are too far apart for its 0.12 threshold. Only worth turning on at 30+ steps. Needs comfyui-minimax-h3-blockcache-T8. blockcache_threshold controls how different two consecutive steps must be before reuse stops; it only means anything with blockcache ON.

And the note the description leads with: stacking Spectrum + TeaCache is no faster than Spectrum alone - they overlap, so pick one rather than enabling both and eating two image-drift penalties for one speed-up.

How it works

Under the hood it's a runtime adapter. When you flip a switch, the node looks up that accelerator's class in ComfyUI's loaded node registry, calls it on your model with the right arguments, and returns the boosted model. If the pack isn't installed, or is installed but its API has drifted from what this node expects, it prints an install/update link to the console and returns your model unchanged - normal speed, no crash. That fail-safe design is why the node is safe to leave in a graph with everything off: it's a pass-through until you opt in.

The single output is your model, boosted. Wire it into the sampler exactly where you'd wire the loader's model output.

Installing it

The node itself ships with the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/KursatAs/ComfyUI-H3-Multishot-Advance

The boosters' packs are genuinely optional - that's the entire point of the node. EasyCache needs nothing. For the others, install them the usual way (ComfyUI Manager or git clone) and restart ComfyUI; the node will detect them at startup.

Common issues

There are almost none, because the design routes around them. If a booster "does nothing," check the console: the pack prints a clear message when the target pack is missing or its API no longer matches, naming the install URL. If output looks subtly different with a booster on, that's the expected behavior, not a bug - the drift is the price of the speed, which is exactly why the description tells you to A/B your own scene first. And if you're on 14 steps, save yourself the confusion and leave blockcache off; it cannot fire there.