Load FastH3 Patch

What this node is actually for

Load FastH3 Patch is the dumb half of a two-node pack, and that's a compliment. It reads one file out of ComfyUI/models/model_patches/ and passes it along the wire as a FASTH3_PATCH. Nothing is applied, nothing goes near your VRAM - it's just "here is the file, the next node does the surgery."

You'd reach for it because FastVideo's FastH3 - the 8-step distillation of MiniMax-H3 - ships as a 44 GB checkpoint, and the entire pitch of this pack is that you don't need it. FastH3 is weight-for-weight the fl2va model, with only the timestep path and 50 extra gate layers genuinely changed. This node is where you pick which of those changes you're willing to download.

Three files, one choice

Don't grab all three. The pack's author splits the delta on purpose:

• ..._patch_full.safetensors - 4.0 GB. The default, and what the pack's example workflow loads, because that workflow runs sparse attention with method=vsa and VSA is the only thing that reads the gate layers.
• ..._patch_adaln.safetensors - 149 MB. The 8-step distillation itself. On a dense graph this is the whole speed-up for roughly LoRA-sized money.
• ..._patch_vsa.safetensors - 3.85 GB. The 50 to_gate_compress layers. Read only by VSA sparse attention; dead weight otherwise.

The full file is the other two combined, so take full or the pair - never all three.

How it reads the file

Inside, it's one line: comfy.utils.load_torch_file(path, safe_load=True, return_metadata=True), and the state dict, metadata and filename go out as a small bundle on the custom FASTH3_PATCH type. Two consequences worth internalising:

• The files are plain safetensors holding FastH3's own tensors verbatim - no quantisation, no factorisation, nothing reconstructed. The apply step routes tensors by what they are, not by filename, so you can rename the file to whatever.safetensors and it still works.
• The whole file is read into system RAM on load. 149 MB is nothing. The 4 GB full patch is 4 GB of RAM before it ever touches the GPU.

The one input and the one output

name is a combo listing whatever sits in models/model_patches/; its tooltip is just "The patch file to read." The output is a single FASTH3_PATCH, which wires into Apply FastH3 Patch's fasth3_patch input and nowhere else. It is not a MODEL and cannot feed a sampler.

So why is loading a separate node at all? Reuse. One load, one or two Apply nodes - handy if you patch two branches of a graph, or chain the split adaLN and VSA files.

Install

ComfyUI Manager → search FastH3 Patcher → install → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Nynxz/ComfyUI-FastH3Patcher

There are no dependencies whatsoever - pyproject.toml declares none, the README says the same, and the pack is pure Python over safetensors and torch, both of which ComfyUI already ships. In an ecosystem where half of what you install drags in a pinned onnxruntime, that's a small miracle. (Same author as the Ideogram helper nodes, so this isn't a first-day GitHub account dropping binaries on you.)

Then put the folder and the file where the combo can see them:

mkdir -p ComfyUI/models/model_patches

Where people get burned

The empty dropdown is the classic. model_patches is a folder ComfyUI core registers but may not have created for you, so a fresh install can boot with no such directory and the combo shows nothing at all. Make the folder, drop the file in, restart. There's a 2025 thread of exactly this confusing people over the USO projector file - same folder, same dead end. It is not models/loras, and it is not models/diffusion_models; patch files never appear in a checkpoint loader.

If the file appears but the run errors, that's the next node's problem, not this one - Apply FastH3 Patch is where the matching happens. And one thing no loader can fix: your base H3 weights are still under MiniMax's community licence, which geofences the US, EU, UK and Korea. Downloading a patch doesn't route around that.