H3 Depth Skip

Not every transformer block in H3 earns its compute on every clip. Some bands of depth are measurably idle - on the released checkpoints, roughly blocks 3 through 17. H3 Depth Skip is the node that drops the named blocks from the forward pass, worth about 1.4× on that band, and it comes with an honest tooltip: read the block contributions off your own model before you trust anyone's band.

How it works

model passes through, patched, and comes back out alongside a report. The inputs:

• blocks - a string like "3-17", "3-17, 22" or "10". Ranges are inclusive. Empty changes nothing - an exact no-op, so it's safe to leave wired-but-blank. The README's guidance is to read them off MiniMax H3 Block Contribution rather than guess, and it notes the quiet band is identical across the fl2va and ref2va releases.
• start_percent / end_percent - where the skip switches on and off (default 0 to 1, always-on). Ending early is the useful setting: skip the quiet band while structure is being decided, then run the full stack for refinement, where the dropped blocks may matter more than their average contribution suggests.
• mode - attention_only (default) keeps the MLP; whole_block drops everything. The default is the better-evidenced choice: attention is ~83% of block time at these sequence lengths, so attention_only captures most of the saving, and the MLP's gate varies only 8× across depth against attention's 68× - much less obviously idle in any band. Reach for whole_block once an A/B says the band really is inert.

The report tells you which blocks were skipped, what fraction of the forward that removes, and what the number does and doesn't mean - worth a glance before you start trusting the speedup.

The honest take

This is a "confirm against your own output" node, not a set-and-forget one. The 3–17 band is a measured property of the released checkpoints, but it's still a hypothesis about your content. The safe recipe: run the A/B with attention_only first (keeps the MLP, so it degrades gracefully), and only switch to whole_block when the result is provably fine. It stacks with H3 Block Cache - different axes of the same save.

Install

Pack install: ComfyUI Manager → "Nynxz H3", or:

cd ComfyUI/custom_nodes
git clone https://github.com/Nynxz/ComfyUI-NynxzH3

Restart. Pure Python, no dependencies, ComfyUI 0.30.0+.

The one-liner

"3-17", attention_only, and a start_percent/end_percent window that lets refinement run full. If your clip starts losing structure late, it's the skipped blocks talking - widen the window or drop the node before you blame the sampler.