H3 Detail Schedule

H3's default shift of 12 spends almost all its steps deciding shape and motion, and exactly one step below sigma 0.5 - the region where fine detail gets resolved. H3 Detail Schedule fixes that by budgeting the refinement phase separately, while keeping every sigma on the exact same shift curve. You wire it into SamplerCustomAdvanced in place of a scheduler, and your detail stops being a single jump.

How it works

Instead of one step budget, you get two phases split at a sigma value:

• structure_steps (default 15) - steps above the split, where global shape and motion are decided. H3 wants a long structure phase; this is what the default shift buys.
• detail_steps (default 8) - steps below the split, the refinement region. Plain shift-12 at 20 steps gives you one of these. That's the whole reason this node exists.
• split (default 0.5) - the sigma where structure ends and refinement begins. Lower spends more of the budget on structure; 0.5 is where H3's own curve has already spent 92% of its range.
• shift (default 12) - the flow shift, matching MiniMax H3 Sigma Shift. The curve is untouched - every sigma emitted lies exactly on it, only sampled more densely at the bottom. Change it here and on the model.
• detail_spacing (default 0) - evens out the refinement steps. At 0, the first refinement step is 3.7× shorter than the structure step before it, which makes res_multistep's second-order term stop contributing for one step at the handoff. At 1, every refinement step is the same size. The split and the final sigma don't move at any setting, and it costs no steps.

Outputs are SIGMAS (into SamplerCustomAdvanced) and a report telling you where the steps landed and the size of the final hop - the number that says whether detail gets resolved or invented.

Why the readout matters

Measured in log-SNR - the coordinate where a flow model's difficulty is actually uniform - plain shift-12 at 23 steps reads out at λ 0.61, while a 15/8 split reads out at λ 2.15: a nat and a half deeper into high SNR. That's the whole pitch: same curve, same trajectory family, but the model gets a real chance to resolve fine detail instead of landing on it.

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

Start at 15/8 with detail_spacing 0 and read the report's final-hop size - that's the honest quality bar. If you want a touch more, that's what the node's sibling, H3 Scheduled Sampler, exists for; this one is the plain, composable version.