MiniMax H3 fixes faces with a latent upscale partway through the run, which
means splitting the generation: some steps, the upscaler, then the rest. The
sampler could only ever run a schedule end to end, so getting the upscaler in
meant dropping shader noise out of the workflow entirely.
Four inputs, the same ones KSampler (Advanced) has: add_noise, start_at_step,
end_at_step and return_with_leftover_noise. They land at the tail of the
optional block, because ComfyUI maps saved widget values by position.
The machinery was already there. The pipeline builds one sigma schedule and
samples it in segments, slicing sigmas[start:end + 1] for each; a window is
that same slicing applied once more at the outer edges. Indices stay absolute,
so the per-segment seed still comes out as seed + start -- which means a split
run draws the same ancestral and SDE noise as the unsplit one it came from.
Decisions worth naming:
- Stages spread across the steps the node actually samples, not the whole
schedule. Two stages over a three-step window are two stages in those three
steps; spread over the schedule, one would land outside the window and never
fire. merge_boundaries takes the window's own edges for this, and its first
parameter is renamed end_step to say so -- measuring the tail against the
schedule length instead would leave a 1-step tail inside a long window.
- stage_progression keeps measuring position in the whole trajectory, so a node
running the last three steps of seven gets the fine end of coarse_to_fine
rather than a fresh coarse-to-fine sweep of its own.
- With add_noise off the opening boundary is not painted. The latent already
carries its noise from whatever ran before, and the shader on a zero tensor
would add back exactly what was turned off. Later boundaries still paint,
since their noise is recovered from the latent rather than made here. The
refusal for an unpaintable latent weighs only the stages that will actually
be painted, so it no longer rejects a latent on behalf of noise nobody makes.
Two things fixed on the way:
- force_full_denoise was dead. KSampler.sample only reads it alongside
last_step, and this code passes sigmas= with last_step unset, so it was
discarded on every call. The window's zeroed sigma tail is what does that job
now. The golden fixtures recorded the flag, so they are re-captured; every
sigma, noise tensor, seed and output across all thirteen is unchanged, which
was checked before blessing them.
- The progress bar counted the schedule rather than the steps being run.
ProgressBar takes its total from the callback, so a windowed run would have
started part-filled and stopped short of the end. At the defaults the numbers
are identical to before.
A schedule with fewer than two sigmas now returns the latent instead of being
handed to the sampler, which is what denoise 0.0 builds.
Verified against comfy.samplers.KSampler itself across seven start/end/leftover
combinations, requiring the sigmas to match bit for bit -- the parity that lets
this node sit on one side of a split and a stock KSampler (Advanced) on the
other. Two golden fixtures cover the two halves. Also run for real on H3:
4 steps, latent upscale, 3 steps, 896x896x124 with audio, in the new
example_workflows/MiniMaxH3_Split_Upscale_SNK_Direct.json.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The Direct node's shader_type list offered three of the four generators this
repository ships. temporal_coherent was registered, covered by the test suite
and reachable from both the standard and legacy dispatch paths -- but absent
from the combo, so no workflow could select it. Nothing tied the advertised
list to the registry, so nothing caught it.
It is the generator best suited to video: 4D simplex with time as a real axis
rather than a per-frame redraw. On real MiniMax H3 weights it degrades far more
gracefully than the others (faint background striping at strength 0.35, where
domain_warp already shows blocking and curl_noise has heavy colour casts), and
it is the only type whose audio level stays flat as strength rises: -24.0 /
-23.7 / -23.7 dB against a -22.6 dB baseline, where domain_warp climbs to -17.0.
Two tests now pin the dropdown and the registry to each other in both
directions. The second fails on the previous list, naming temporal_coherent.
The tooltips implied the whole 0.0-1.0 strength range was usable. Measured on
H3 at 608x352 and 8 steps it is not:
domain_warp + multiply photoreal to ~0.25, gone by 0.75
tensor_field tolerates ~0.5
curl_noise ~0.2, heavy colour casts by 0.5
soft_light / difference clean at 0.3, the gentlest blends
add ~0.15, the most aggressive
shape masks ~0.2; at 0.6 the mask is drawn into the picture
use_temporal_coherence ~0.2; 0.5 swamps the frame
use_temporal_coherence was the worst offender: it claimed to help "maintain
frame-to-frame consistency", when holding one seed across frames stops the
pattern varying between them and so reinforces it rather than averaging it out.
The tooltips now carry these ceilings, rank the blend modes by aggressiveness,
and note that the live preview only draws three of the four types.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The Direct node now dispatches: "standard" (default) runs the new pipeline,
"legacy" runs the frozen pre-2.0 one. Measured through the node with a
recording sampler, 20 steps:
standard, 1 stage 1 call x 20 steps, from sigma 14.61
standard, 2 stages 2 calls x 10 steps, from 14.61 then 1.48
standard, denoise 0.6 1 call x 20 steps, from 2.23
standard, 3 injections 4 calls x 5 steps, 14.61 / 3.87 / 1.48 / 0.60
legacy, 2 stages 2 calls, each rebuilding a full schedule
The 1.48 is the fix: stages continue one trajectory instead of restarting at
maximum noise, where flow models discard the previous stage entirely. The 2.23
is denoise finally reaching the schedule.
Also on the node:
- sequential_distribution, injection_distribution and fast_high_channel_noise
become real optional inputs. As V1 `hidden` tuple inputs ComfyUI never
delivered them, so they were stuck at their defaults.
- the debug/visualisation hidden inputs are gone; they drove stub no-ops.
- IS_CHANGED is removed: it only restated widget values that are already part
of the cache key, and would have rejected the new input.
- new widgets are appended last, so saved workflows keep their widget order.
web/src/sampling_mode_migration.ts switches nodes loaded from pre-2.0
workflows to "legacy", recognising them by the absence of the snk_version
property, so existing seeds keep reproducing.
186 Python tests and 85 web tests pass; the 11 legacy goldens now exercise the
legacy branch through this dispatch.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
