H3 Relay · Attention Backend

The H3 Relay · Attention Backend node is a one-input, one-output patch: it takes your loaded H3_RELAY_MODEL bundle, swaps in a different optimized attention implementation, and hands you back the same bundle with the attention patched. That's it. There's no generation here, no settings beyond a dropdown, no video output. It exists because MiniMax H3 is a 33B omni-modal model and the attention kernel you pick is the difference between "H3 runs at a usable speed" and "H3 crawls."

If you've been around ComfyUI a while, this is the same trick as patching set_model_optimized_attention on any diffusion model - a model-level swap that replaces the default attention path without touching weights. The bundle it operates on also carries an internal cache fingerprint, and the node updates that fingerprint whenever it patches. So changing the attention backend invalidates derived cached artifacts, but only derived ones - your raw H3 shots stay valid.

The dropdown is honest, which is rarer than it should be

Here's the part I genuinely like about this node: the choices it shows you are filtered by what's actually registered in your installed ComfyUI. It asks the installed environment which backends are really available, and only lists those. The source registers four labels - Comfy Kitchen, PyTorch, Sage, and Sage 3 - but the drop-down you see will only contain the ones that exist in your setup.

• Comfy Kitchen - the reference H3 graph from the pack uses this. It's the safe default if it's present.
• pytorch attention - stock, always available, the fallback that works everywhere but isn't optimized.
• sage attention - only shown when the optional Sage package/backend is actually installed. H3 Relay does not silently pretend Sage is there.

That honesty is the feature. Plenty of packs draw a dropdown full of options and then quietly fall back to whatever works; this one refuses to show you a backend you don't have. If you don't see Sage, that's not a bug - it means Sage isn't installed.

What a beginner should actually do

Load the model with either H3 Relay · H3 Model Loader or the Hybrid Loader, run the bundle through this node, then send it to Generate Shot. Try PyTorch first if you're not sure anything else is installed - it's the baseline. If Comfy Kitchen shows up, that's the one the reference workflow uses, so it's a reasonable first optimization. Don't chase Sage unless you know you have it.

One real-world note: because this is a model patch, it has nothing to do with quality. Same seed, same prompt, same scheduler - you'll get near-identical results at different speeds. If a shot looks different after switching backends, something else in your graph changed, or you've hit a numerical edge case. Don't rebuild a whole workflow because of this node.

Install and context

ComfyUI Manager search H3 Relay, or git clone https://github.com/akatz-ai/h3-relay.git into custom_nodes and restart. Requires ComfyUI 0.32.0+, FFmpeg on PATH, and the H3 model files from MODELS.md - the pack never redistributes weights, and the H3 community license excludes US/EU/UK/Korea, so check that before you go far down this road. This node itself has zero extra Python dependencies; the whole pack installs clean.