凤希AI - MiniMax块缓存加速

FxAiMiniMaxBlockCache (凤希AI - MiniMax块缓存加速) is the fxai-toolkit speed lever for MiniMax H3, and it's the same family of trick as TeaCache / block-cache: during early sampling steps, cache the outputs of the model's DiT blocks; on later steps, skip recomputing blocks whose output would barely change and reuse the cached result. The result is fewer forward passes per step - less VRAM churn and a faster render - at the cost of some accuracy, which the node exposes honestly as a four-stop tradeoff dial instead of hiding it.

The dial is 速度档位 and the choices spell out the trade directly:

• 不加速 - official original speed, no caching.
• 画质优先 - conservative caching, smallest error.
• 均衡 (default) - the middle ground the author recommends.
• 极速 - fastest, largest error margin.

It also reads your 采样步数 and behaves more conservatively at low step counts, because with fewer steps to amortize the cache over, a mistake costs more of the render. The node finds the MiniMaxH3 model inside whatever you pass it, clones the model, and installs a patch - and this only works on actual MiniMax H3 models, so feeding it anything else is a hard error (FxAiMiniMaxBlockCache 仅支持 MiniMax H3 模型). Outputs are 模型 (the patched model - wire it into your sampler) and 采样步数 (the same value passed through, so you can drive the sampler's step count from one place).

What it leans on: this is where the pack's core patch matters. fxai-toolkit installs MiniMaxH3Patch at import time (via fxai_minimax_core_patch) that adds a ("block_loop", 0) hook to the H3 model so the cache can intercept block execution - plus a fix for an official ComfyUI H3 bug where reference images overwrote the first/last-frame keyframe conditioning. What that means for you: the node only makes sense on a ComfyUI build with native MiniMaxH3 support, and the pack's patch has to load, so update the pack alongside ComfyUI core.

Install: ComfyUI Manager → search "fxai-toolkit", or cd ComfyUI/custom_nodes && git clone https://github.com/fxai666/fxai-toolkit, restart. Auto-installs soundfile/psutil, wants FFmpeg, all labels Chinese (凤希AI, QQ group 775649071, Bilibili).

The honest take: block caching is a real technique (it's how a lot of local diffusion acceleration works), but "极速" is for previews and draft passes, not finals - if you're rendering something you care about, run 画质优先 or skip it. And because the cache behavior depends on the model's block layout, an H3 variant that renames or resizes its DiT blocks can silently fall back or error; if you see a sudden slowdown after a model update, that's the cache quietly bailing out, which is the safe failure mode. It's the node you reach for when a long H3 loop is dragging and you'll take "visibly faster, marginally softer" over "perfect and slow."