ApplyMSWMSAAttentionSimple

If ApplyMSWMSAAttention is the advanced SD1.5 speed hack, this is the version you actually reach for on a Tuesday. Same MSW-MSA windowed-attention trick from HiDiffusion, same ~30% faster high-res generations on SD1.5, zero block-number homework. It picks the attention blocks for you based on your model type and applies the effect from 20% to 100% of the run - exactly the settings the author has landed on through testing.

It's the perfect first stop: wire your model in, pick auto, forget about it. If you later want the manual controls (which blocks get patched, when the effect starts and ends), the advanced node exists, and the simple node even tells you how to replicate its behavior there.

Inputs and output

Only three things on the face of this node:

• model_type - auto, SD15, or SDXL. auto inspects the model's latent format and guesses, which is genuinely reliable in practice. Pick SD15 explicitly if you're on SD 1.4 or 2.x.
• model - the MODEL to patch.
• Output: a single patched MODEL, straight into your KSampler.

That's the entire node. auto + model in, model out. The SD15 preset patches input blocks 1,2 and output blocks 11,10,9; SDXL uses 4,5 in and 3,4,5 out. If you want a reproducible generation, watch the log when it runs - the node prints its preset to the console, so you can copy those block numbers and timings into the advanced node and tweak from there.

Install

Same story as the rest of the pack. ComfyUI Manager → search comfyui_jankhidiffusion → Install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/blepping/comfyui_jankhidiffusion

No model files, no requirements.txt, nothing extra to download. The pack is pure Python against the ComfyUI API, which is exactly how we like our custom nodes.

Caveats worth knowing

This is a wrapper, so every caveat from the advanced node applies: it only supports SD1.x, SD2.x and SDXL, and it's really an SD1.5 play - the author is upfront that SDXL barely benefits from MSW-MSA. Keep resolutions multiples of 64/128 to avoid tensor size mismatches. And don't stack it with other attention patches that hit the same blocks - HyperTile, Deep Cache, nearsighted attention - or they'll fight each other and you'll get the "Incompatible model patches or bad resolution" warning in your log.

For SD1.5 users generating at 1024×1024 and up, this is a five-second install for a genuinely large speed bump. Add ApplyRAUNetSimple from the same pack when you want to push the resolution itself higher, and you've got a two-node high-res setup that punches way above its weight.