MiniMax H3 Attention Lab [Experimental]

MiniMax's closed H3 generation reportedly leans on a sparse-attention topology, and if you've run H3 locally on a 33B model you've probably wished for the cheaper attention too. MiniMax H3 Attention Lab is this pack's attempt to poke at that - but read the label before you get excited: this node is not an implementation of MiniMax's unreleased sparse topology, and the README is explicit that it should not be treated as a production acceleration path. It's a guarded research and diagnostics sandbox for H3's packed attention layout.

The part that's genuinely useful rather than aspirational is diagnostic mode. H3 packs text, conditioning/reference, audio, and video into one sequence, and the video portion lives on a (1, 2, 2) DiT patch grid over the 16×-downsampled latent. Diagnostic mode inspects selected transformer blocks and reports the layout - how many rows each segment (text, reference, audio, video) occupies, the sequence length, the signature - without touching behavior. If you're building something that cares about H3's packing (which is half of why this pack exists), that visibility is worth having.

experimental_sparse is the actual sparse-attention experiment: a local-window mask for video-to-video keys (spatially local on the 2×2 patch grid, while all temporal positions in the local spatial column stay visible, and text/reference/audio keys stay global), on the layers you choose. It's structurally guarded - mode=native returns your model unchanged, and the code validates inputs hard - but the author's own framing is that sparse attention here is investigative, not a reconstruction of MiniMax's proprietary design and not a vetted speed path.

Inputs you'd touch in diagnostic/experimental mode: mode (native default, diagnostic, experimental_sparse), layers - a comma-separated string of transformer block indices, default "8,16,24,32,40", which gets parsed and validated (empty or non-integer lists are rejected), sparse_window (local radius, default 4, 1–32), global_heads (how many heads keep global attention, default 8, 0–56), and max_sequence (cap on sequence length, default 8192). Outputs: the (possibly patched) model plus a metrics object you can sink into Metrics JSON. In native mode it's a straight passthrough.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/xmarre/MiniMax-H3-Flow-Aligned-Regenerate.git

Restart. No extra Python dependencies; the hard requirement is that your ComfyUI exposes the native H3 24/32-channel, (1,2,2), shift-12/3 contract - otherwise the pack fails closed rather than pretending.

The honest take

Run this if you're researching H3's attention layout or testing whether a local-window scheme could survive on your hardware. Don't wire experimental_sparse into a production render expecting a free speedup - the author hasn't validated it as one, it's explicitly not the closed model's topology, and quality gates here are decoded media you'll have to judge yourself. This whole pack is a single-author research effort with zero community footprint so far, which means the only peer review these claims have gotten is the author's own test suite. Treat the lab as a lab.