MiniMax H3 Aicg 加速设置

A 33B video model with a native audio stream is not going to feel fast on consumer hardware. Everyone who runs H3 locally ends up in the same two places: attention kernels, and skipping denoise steps. This node (MiniMax H3 Aicg 加速设置) is the pack's dashboard for both, and its own description gives you the wiring order in one line: loader → this node → generation/render node.

It takes h3_bundle and returns h3_bundle. That's it. It's a configuration node that patches the bundle in place.

Attention: the boring-but-real win

注意力加速(Sage) defaults to 自动（Sage 优先）, which means it looks for SageAttention in ComfyUI's attention-backend registry and uses it if present - and if you don't have it installed, quietly leaves ComfyUI's current backend alone. Your manual options are 关闭, SageAttention, PyTorch SDPA, Sub-Quadratic and Split. It reuses whatever ComfyUI already knows about, so you're not installing a second attention shim just for this pack.

Is Sage a free lunch? Mostly, and with an asterisk. It's the standard first lever for video models for a reason, and it's also the cause of the most entertaining class of bug reports in the ecosystem - patchy output, matrix-code-looking artefacts, occasionally black frames - which is exactly why this node exists rather than a config file. When something looks wrong, set this to 关闭 and re-run before you blame the model. That one test separates two completely different afternoons of debugging.

MotionCache: real speed, real cost

运动缓存(MotionCache) is the interesting bit. The mechanism, from the source: at each step it compares the incoming latent against the previous step's, scales the change by MC运动强度, accumulates it, and - if the accumulated change is still under MC复用阈值 and it hasn't already skipped too many in a row - reuses the previous step's residual instead of running the transformer. You get the step "for free." Skipping is bounded by MC最大连跳, gated to a window between MC起始% and MC结束%, protected by MC预热步数 of always-computed steps at the start, and the change is measured on a subsampled tensor (MC采样间隔) so the measurement itself isn't the bottleneck.

In other words it's a temporal heuristic: if consecutive denoise steps are barely moving the latent, assume they'll keep barely moving it. That assumption holds on slow, locked-off shots and breaks the moment something actually moves.

The knobs, in the order I'd touch them:

• MC复用阈值 (0.08 by default) - raise it and it skips more. This is the aggressiveness dial.
• MC运动强度 (0.1) - the inverse guardrail; raise it and motion makes the node bail out of reuse sooner.
• MC预热步数 (4) - steps that always compute. Don't zero this; the skeleton of the frame is built early.
• MC最大连跳 (2) - hard cap on consecutive skips. The single best protection against the whole clip drifting.
• MC起始% / MC结束% (0.15 / 0.95) - the window. The tail matters: the last steps are what crispen detail, so keep MC结束% below 1.0.
• MC采样间隔 (8) - measurement stride. Higher = cheaper comparison, coarser judgement.

How I'd actually use it

MotionCache on for exploration, off for the final take. It's a draft-speed tool: you're paying for it with detail in the exact places people look at (edges, skin, fine texture), and a skipped step cannot be recovered by re-running. Set the seed, find the take you want with cache on, then flip it off and re-run that seed.

Attention acceleration stays on, in 自动, unless you have a reason.

Installing

Same pack, same install: ComfyUI Manager → search AICG3D or 短剧工作台, or:

cd ComfyUI/custom_nodes
git clone https://github.com/JGRFW/comfyui-AICG3D

Restart, then find it under AICG3D. The node itself needs nothing extra; SageAttention is a separate install and entirely optional - that's what 自动 is for.

Where people get burned

If you've been following this pack for a while: MotionCache used to live on the loader, and it was removed from there. The comment in the loader source is blunt about why - the loader keeps exactly one acceleration widget (注意力加速) so old workflows don't shift their widget positions, and the MotionCache params moved here. So a workflow from before that change has nowhere to put its cache settings; swap in this node.

Also: node-level caching. These values are part of the bundle, so changing them does change what runs - but if you're hunting a speed difference and seeing none, check you're actually re-executing the sampler and not loading a cached result from an unchanged graph.