MiniMax H3 RefDelta Solver

ComfyUI MiniMax-H3 RefDelta Solver

A dedicated RefDelta-aware sampler family for the MiniMax-H3 Pruned Ref-Delta Fused rank-1024 checkpoint.

The Stability Sampler can now use ER-SDE, SEEDS-2, SEEDS-3, SA-Solver PEC, or SA-Solver PECE from one base_sampler dropdown. ER-SDE remains the saved-workflow and production default until the additional backends complete the same real-media validation gate.

The target checkpoint is conceptually FL2VA plus a rank-1024 approximation of the Ref2VA-from-FL2VA parameter delta, with exact/non-matrix delta pieces where applicable, and the production checkpoint also contains INT8 ConvRot approximation effects. Genuine Ref2VA is useful for same-state diagnostics of the imported delta. It is not a quality oracle and is not used to drive the production scheduler.

Production recommendation

Use:

• MiniMax H3 RefDelta Stability Sampler;
• base_sampler = er_sde for the currently validated production default, or select seeds_2, seeds_3, sa_solver, or sa_solver_pece for the corresponding enhanced backend;
• remember that steps means outer sigma intervals: SEEDS-2/3 and active PECE expose substantially more logical H3 model-call opportunities than the same numeric step count on ER-SDE/SA PEC;
• ComfyUI BasicScheduler with scheduler = beta;
• the normal MiniMax-H3 ModelSamplingAV path;
• Spectrum MiniMax H3 v0.2.20+ for the existing ER-SDE interop; RefDelta SEEDS/SA-Solver interoperability requires the companion Spectrum multi-backend interop change (PR #91 or a release containing it). Active RefDelta SA-Solver PECE additionally requires the updated #91 active-PECE composition contract.

The custom RefDelta scheduler is retained only for saved-workflow compatibility and scheduler research. The bundled r1024_provisional profile is neutral and should not be interpreted as a calibrated production schedule.

Install

Clone this repository into ComfyUI/custom_nodes and restart ComfyUI:

cd ComfyUI/custom_nodes
git clone https://github.com/xmarre/ComfyUI-MiniMax-H3-RefDelta-Solver.git

The custom node has no runtime dependencies beyond current ComfyUI.

Nodes

MiniMax H3 RefDelta Stability Sampler

This is the recommended production node. Its production defaults are built directly into the node while every production-relevant knob remains editable. There is no separate sampler profile or calibration file to select. Diagnostic capture controls and legacy controller-mode selection are kept out of this node so ordinary workflows do not mix production tuning with research capture state.

Base sampler family

base_sampler selects the numerical solver underneath the shared RefDelta trajectory and stochastic-stability policy:

| Value | Native solver geometry retained | H3 model-call opportunities for N outer steps* | RefDelta integration | | --- | --- | ---: | --- | | er_sde | ComfyUI ER-SDE stages/noise | N | Existing full RefDelta path, including adaptive ER order | | seeds_2 | Native two-stage SEEDS and correlated stochastic decomposition | 2N - 1 | Outer-trajectory correction plus one frozen stochastic gate reused across both correlated noise segments | | seeds_3 | Native three-stage SEEDS and correlated stochastic decomposition | 3N - 2 | Outer-trajectory correction plus one frozen stochastic gate reused across all three correlated noise segments | | sa_solver | Native SA-Solver PEC Adams predictor/corrector | N | RefDelta correction on each PEC endpoint evaluation plus gated native predictor noise | | sa_solver_pece | Native SA-Solver PECE predicted/corrected topology | 2N - 1 | P0 and exact corrected C_i calls own persistent RefDelta evidence; later predicted P_i calls stay ephemeral current-corrector inputs; native predictor noise is gated from the persistent endpoint |

* Counts above assume the usual terminal-zero sigma schedule. The sa_solver_pece 2N - 1 row additionally assumes active PECE with corrector_order > 0; with corrector_order = 0, it reduces to N logical H3 model-call opportunities. The ComfyUI steps setting counts outer sigma intervals, not logical H3 model-call opportunities. SEEDS-2/3 expose internal-stage calls on every nonterminal outer interval, while active PECE exposes both predicted and corrected calls after P0.

Concrete examples:

The numeric cells below are logical H3 model-call opportunities:

| Outer steps | ER-SDE / SA PEC | SEEDS-2 | SEEDS-3 | SA-Solver PECE | | ---: | ---: | ---: | ---: | ---: | | 10 | 10 | 19 | 28 | 19 | | 19 | 19 | 37 | 55 | 37 |

So a 10-step sa_solver_pece run legitimately exposes 19 logical H3 model-call opportunities (P0 plus nine predicted/corrected pairs), and a 10-step SEEDS-3 run exposes 28. This does not mean the scheduler secretly changed the outer step count. It is native solver geometry, and raw step counts are therefore not NFE-equivalent across these backends.

The SEEDS internal denoiser stages are intentionally not rewritten by RefDelta. Their multi-stage stochastic construction depends on correlated noise over overlapping intervals; the RefDelta gate is resolved once per outer interval and reused for every native segment so stage count cannot advance the controller or alter the solver's correlation structure.

sa_solver means the native PEC path (use_pece = false). sa_solver_pece uses the same current ComfyUI defaults—predictor_order = 3, corrector_order = 4, and simple_order_2 = false—with use_pece = true.

Active PECE is not treated as a two-stage fixed-stride sampler. Native ComfyUI calls P0, P1, C1, P2, C2, ...: the predicted and corrected evaluations at an outer coordinate share sigma but use different latent states. RefDelta mirrors native endpoint replacement:

persistent RefDelta evidence = P0, C1, C2, C3, ...
ephemeral PECE inputs        = P1, P2, P3, ...

Only persistent endpoint calls update RefDelta trajectory derivatives, spatiotemporal stability evidence, and the stochastic gate used for the following predictor noise. This avoids double-counting the same outer coordinate and prevents a Spectrum-predicted P_i from entering persistent RefDelta evidence. Trajectory correction is likewise applied to the persistent endpoint value, not twice to both P_i and C_i.

s_noise is shared by all five backends. max_stage and adaptive_order are ER-SDE-specific numerical-order controls; SEEDS has its own fixed native stage geometry and SA-Solver has its native Adams orders. Calibration capture remains ER-SDE-only because its saved-state format is defined on the original ER-SDE outer trajectory.

With RefDelta controls disabled, each new backend delegates directly to the corresponding native ComfyUI sampler. The instrumented zero-adaptation path is also covered by native same-input/same-noise parity fixtures so the wrapper cannot silently perturb solver behavior.

The additional backends are newly implemented and automated-parity tested. They are not described here as media-validated production replacements for ER-SDE yet.

Current empirically validated defaults:

adaptive_order = true
risk_sensitivity = 1.00

stochastic_adaptation_strength = 0.50
minimum_stochastic_multiplier = 0.50

trajectory_correction = true
video_correction_strength = 0.15
audio_correction_strength = 0.05
correction_bound = 0.50
endpoint_fidelity_fraction = 0.15

s_noise = 1.00
max_stage = 3

video_stochastic_strength_scale = 1.00
audio_stochastic_strength_scale = 1.00
static_video_stochastic_adaptation_strength = 0.25
video_stability_restore_strength = 1.00

video_stability_motion_low = 0.15
video_stability_motion_high = 0.60
video_stability_diffusion_low = 0.05
video_stability_diffusion_high = 0.50
video_stability_diffusion_weight = 0.20
video_stability_normalization_floor = 0.10
video_stability_gamma = 1.00

video_stability_spatial_radius = 2
video_stability_temporal_radius = 2
video_stability_ema = 0.70
video_stability_start_fraction = 0.10
video_stability_full_fraction = 0.30
stochastic_gate_slew_limit = 0.00

These values are the current production candidate for the rank-1024 INT8 ConvRot checkpoint, not a claim that one setting is universally optimal for every prompt/reference stack. The node keeps the knobs exposed specifically so tuning does not require code changes.

Same-seed validation of the streamlined Stability Sampler against the advanced node produced cell-for-cell identical CSV telemetry and byte-for-byte identical JSONL telemetry when both were configured with these values. The streamlined node therefore changes workflow UX, not the sampler/controller path.

What the stability controller does

The dynamic video policy is driven by the actual-only video risk. The audio stream has its own scalar risk/gate. Video additionally receives a smooth [B,1,T,H,W] latent-space stability map built from:

• channel-RMS motion between neighboring latent time slices;
• channel-RMS change from the previous actual denoised prediction;
• configurable temporal/diffusion thresholds and diffusion weighting;
• gamma shaping, spatial/temporal smoothing, actual-only EMA, and a progress ramp.

Static regions interpolate toward static_video_stochastic_adaptation_strength; moving or uncertain regions remain closer to the dynamic stochastic_adaptation_strength policy. The controller does not decode pixels, use optical flow, or add model evaluations/NFE.

The static target is deliberately a separate policy, not a global stochastic floor. In real artifact testing this distinction mattered: globally raising the minimum multiplier could introduce its own jitter, while selective low-motion restoration could suppress fine-detail background shimmer without forcing the whole video onto the same stochastic behavior.

endpoint_fidelity_fraction remains independent and fades the final adapted gate toward native multiplier 1 near the endpoint. Optional stochastic_gate_slew_limit acts before endpoint restoration; 0 disables it.

Main tuning controls

| Control | Effect | | --- | --- | | stochastic_adaptation_strength | Dynamic-region adaptation strength | | minimum_stochastic_multiplier | Global lower bound on the stochastic multiplier | | static_video_stochastic_adaptation_strength | Static-region target policy | | video_stability_motion_low/high | Which latent regions count as temporally stable | | video_stability_diffusion_low/high | Whether those regions are also settled across actual denoising steps | | video_stability_diffusion_weight | Importance of actual-to-actual stability | | video_stability_gamma | Selectivity of restoration confidence | | spatial/temporal radii + EMA | Smoothness/persistence of the stability map | | start/full fractions | When stability restoration becomes active | | stochastic_gate_slew_limit | Optional per-step gate-change bound |

Practical tuning order:

1. Keep stochastic_adaptation_strength, the stochastic floor, and static target separate.
2. If static detail still shimmers, inspect telemetry/maps before widening thresholds.
3. If the map identifies the correct regions but restoration is too weak, tune the static target rather than globally raising the floor.
4. If composition/motion changes too much, make restoration more selective with thresholds or gamma before applying a stronger policy everywhere.
5. Keep audio tuning independent through audio_stochastic_strength_scale.

[Advanced/Diagnostic] MiniMax H3 RefDelta Sampler

This is the original node ID retained for saved workflows, manual controller experiments, and diagnostic capture. Its defaults intentionally preserve released behavior:

• stochastic_control_mode = legacy_global;
• stochastic_adaptation_strength = 0.50;
• minimum_stochastic_multiplier = 0.25;
• trajectory_correction = false.

It additionally exposes:

• legacy_global, streamwise, and spatiotemporal_stability controller selection;
• calibration_capture / calibration_id for disk-backed diagnostic capture;
• the same detailed stability controls as the production node.

The legacy defaults are not the current recommended rank-1024 production defaults; they exist so older saved workflows do not silently change behavior after upgrading.

For a strict stock ER-SDE baseline:

adaptive_order = false
stochastic_adaptation_strength = 0
trajectory_correction = false
debug_telemetry = false
calibration_capture = false

That path delegates directly to current ComfyUI sample_er_sde.

MiniMax H3 RefDelta Comparison Replay

This is the preferred same-state diagnostic path. The historical calibration_* names are kept for workflow/file compatibility, but the data is now treated explicitly as diagnostic comparison data, not scheduler calibration.

A complete comparison uses three executions of the same workflow while loading one full H3 model at a time.

Pass 1 — fused diagnostic capture

Use the advanced/diagnostic sampler with the fused model and enable:

calibration_capture = true
calibration_id = int8-convrot-r1024-test-01

For a clean vector-field comparison, use stock beta, Spectrum off, and optionally disable sampler adaptations:

adaptive_order = false
stochastic_adaptation_strength = 0
trajectory_correction = false

The capture writes exact packed sampler states, fused x0 values, sigmas, actual/forecast classification, scalar telemetry, and the final pre-inverse-scaling sampler state under:

ComfyUI/output/refdelta_calibration/<calibration_id>/

Pass 2 — FL2VA

Switch the MODEL loader to FL2VA, replace the sampler with MiniMax H3 RefDelta Comparison Replay, use the same calibration_id, and set:

comparison_label = fl2va

Only captured actual steps are evaluated at the exact captured state/sigma.

Pass 3 — Ref2VA

Switch the MODEL loader to genuine Ref2VA and set:

comparison_label = ref2va

The completed FL2VA pass is required. Replay measures the true Ref2VA - FL2VA field against the fused fused - FL2VA field, including video/audio cosine, magnitude, projection, residual, and orthogonal components.

Comparison replay always returns the captured fused final sampler state, preserving the original Continuum trajectory across diagnostic passes. Spectrum is rejected on comparison passes. FL2VA/Ref2VA metrics are explanatory and never enter production stochastic control or scheduler density automatically.

MiniMax H3 Uniform Flow Scheduler [Experimental]

This separate scheduler node grew out of the strongest scheduler lead seen in real MiniMax-H3 testing: ComfyUI ddim_uniform produced many quick cuts, coherent action, and unusually high scene/action variety. The released laboratory keeps that exact control but now defaults to the best setting from the completed 19-step MiniMax-H3 + ER-SDE media A/B pass:

steps = 19
mode = phase_offset_uniform
phase = 0.50
denoise = 1.00

The phase sweep was materially informative. 0.60 and 0.65 reintroduced visible flashing; 0.55 removed the flashing but produced weaker action than 0.50; 0.45 was worse again. 0.50 gave the best observed balance of coherent action, stable visuals, useful shot variety, and audio behavior in the tested workflow. Small phase changes also altered spoken intonation, confirming that placement changes the shared AV trajectory rather than acting as a video-only quality knob.

Other scheduler families remain available as research controls. The real-media pass found uniform_linspace solid; neutral power_uniform and neutral piecewise_structure_refinement matched it as expected. uniform_refinement_tail and trailing_refined could produce very good results, but the refinement-tail family was more inventive/unstable across repeated tests. av_arc_length produced strong, varied shots but also an artifacted shot in the tested pass. Default asymmetric_beta(0.6, 0.6) was unstable, with rapid shot repetition and doubled audio. These observations are empirical results from the tested H3 workflow, not universal rankings for every prompt, checkpoint, or step count.

Connect the node's SIGMAS output to the same custom sampling path and continue using ER-SDE. “DDIM” here describes time-point spacing; this node does not use a DDIM sampler or replace ER-SDE. The current experimental recommendation is the default phase_offset_uniform / 0.50 / 19 steps preset. The conservative production fallback remains the RefDelta Stability Sampler with stock BasicScheduler(beta) until broader cross-prompt and cross-seed validation justifies changing the global recommendation.

legacy_ddim_uniform delegates to ComfyUI's own scheduler and reproduces the full BasicScheduler behavior, including table index 1, integer floor stride, reversal, occasional extra point, and final tail slice. This unusual placement remains the exact high-variety control. Exact parity is tested against the pinned ComfyUI revisions in CI.

All other modes choose one descending trajectory on H3's shared base flow coordinate u and map it through the loaded model's video shift. H3's ModelSamplingAV and model code remain authoritative for the corresponding audio time. The node reads shift and audio_shift from the loaded model; it never constructs independent audio/video clocks.

Available modes:

| Mode | Definition / experiment | | --- | --- | | legacy_ddim_uniform | Exact current ComfyUI control, including integer table-stride aliasing | | uniform_linspace | Inclusive u=1..0 linspace with steps+1 points | | phase_offset_uniform | Nonterminal u_i = 1 - (i + phase) / steps, then exact zero; released default is phase=0.50 | | power_uniform | u(x)=(1-x)^power; power=1 is exactly uniform_linspace | | uniform_refinement_tail | Uniform body from u=1 to explicit tail_start, then a power-refined tail using tail_steps | | trailing_refined | Uses Diffusers-style final training-index normalization (N-1)/N, joins at tail_start, and explicitly refines to zero | | asymmetric_beta | Continuous beta quantiles in shared base time; deliberately distinct from stock rounded-table beta | | av_arc_length | Blends uniform placement with equal joint video/audio shifted-flow arc length | | piecewise_structure_refinement | Continuous two-segment progress warp with independent structure/detail powers | | curvature_profile | Immutable offline production-telemetry density; bundled h3_uniform_neutral is neutral |

Controls are marked advanced where appropriate and affect only their named modes. In particular, phase, power, tail controls, beta parameters, arc controls, piecewise controls, profile, and profile_path are inert outside their respective mode. auto_tail_steps=true selects min(5, effective_steps - 1) after denoise expansion; disable it to use the explicit tail_steps value. An explicit tail_steps=0 remains the exact uniform-baseline identity.

Continuous modes whose definition starts at u=1 deliberately retain that endpoint. Current ER-SDE applies ComfyUI's offset_first_sigma_for_snr safety adjustment to such a first point for a CONST flow model; subsequent points are unchanged. Phase-offset, positive-tail trailing_refined, and legacy_ddim_uniform intentionally start below one.

denoise < 1 follows BasicScheduler semantics: the node constructs the corresponding longer schedule and returns its exact tail. Outputs are deterministic CPU float32 tensors, strictly decreasing, finite, and terminated by exact zero. Invalid H3 metadata, duplicate points, impossible tails, and legacy step counts beyond the table's unique capacity fail explicitly.

Schedule curves and implied audio times can be exported without running the model:

python tools/inspect_h3_schedule.py --mode phase_offset_uniform --steps 19 --phase 0.50 --format csv

MiniMax H3 SA-Solver Scheduler [Experimental]

This separate node investigates point placement for native ComfyUI SA-Solver PEC and PECE. It changes only the outer sigma coordinates. SA-Solver remains authoritative for predictor/corrector order, PECE endpoint replacement, stochastic variance, RNG/noise draws, callbacks, and model-call count.

The default remains:

steps = 10
mode = simple_control
denoise = 1.0

simple_control delegates to current ComfyUI simple and preserves exact BasicScheduler-style denoise-tail behavior.

Final 10-outer-step production testing covered the real MiniMax-H3 stack with RefDelta SA-Solver PECE, Spectrum, DiffAid, Untwist-RoPE, and H3 Continuum. Both public modes produced acceptable decoded output. simple_control was slightly preferred perceptually over simple_adams_bounded, so the exact ComfyUI-simple control remains the default. The bounded mode remains available as an experimental research alternative; its lower worst native Adams L1 coefficient norm is not presented as a decoded-media quality win.

The companion Spectrum active-PECE tests also found the balanced forecast policy perceptually preferable to max_speed in the tested workflow while both remained structurally clean. Spectrum owns that forecast cadence; this scheduler only supplies the outer sigma coordinates.

Why the original lambda-uniform design was rejected

SA-Solver constructs its Adams coefficients in half-log-SNR/lambda space. For MiniMax-H3 CONST flow sampling, ComfyUI uses effectively:

lambda = log((1 - sigma) / sigma)

That fact does not imply that scheduler nodes should be uniformly spaced in lambda. The first PR implementation tested that hypothesis with simple_lambda_uniform and simple_lambda_blend. Both produced visibly bugged decoded MiniMax-H3 output in the real production workflow and therefore failed the media gate.

The numerical failure mode was clear in shared H3 base time. At 10 steps, full lambda-uniform spacing compressed the first control interval from about 0.1 to 0.000355 while later intervals grew to roughly 0.30. The 0.50 blend still compressed the first interval to about 0.00625 and expanded another beyond 0.20. The Adams coefficients remained finite, so finite coefficient geometry alone was not a sufficient media-validity test.

Those failed modes have been removed from the user-facing node. The inspection tool keeps them only as explicitly named diagnostic controls:

• failed_simple_lambda_uniform;
• failed_simple_lambda_blend.

Bounded Adams-conditioned replacement

The public research candidate is now:

simple_adams_bounded

It treats the ComfyUI simple/shared-base trajectory as primary geometry and uses native SA coefficient information only to make one tightly bounded local adjustment.

For the longer full schedule implied by steps and denoise:

1. build exact ComfyUI simple;
2. convert that schedule to H3 shared base time;
3. compute native predictor-order-3/corrector-order-4 SA coefficients using ComfyUI's own compute_stochastic_adams_b_coeffs;
4. identify the record with the largest coefficient L1 norm;
5. consider only the interpolation-support nodes used by that record;
6. move at most one interior base-time node, transferring width only between its two adjacent intervals;
7. search the deterministic local displacements ±0.5 and ±1.0 times a 12.4% local-interval guard;
8. select the trial with the lowest global worst Adams L1 norm;
9. keep simple unchanged if no trial improves that objective.

The 12.4% internal search guard exists so float32 sigma round-tripping remains safely inside the public hard contract:

0.875 * Δu_simple_i <= Δu_candidate_i <= 1.125 * Δu_simple_i
max |u_candidate_i - u_simple_i| <= 0.125 * mean(Δu_simple)

The contract is checked after conversion back to the actual float32 sigma schedule that downstream samplers receive. Invalid or non-monotone candidates are rejected rather than silently clipped into shape.

The candidate remains deliberately modest:

| Outer steps | Control node -> bounded node | Min/max interval ratio vs simple | Worst Adams L1: simple -> bounded | | ---: | --- | --- | --- | | 10 | u[4] ≈ 0.600000 -> 0.612400 | 0.876 / 1.124 | 0.426531 -> 0.402478 | | 19 | u[14] ≈ 0.264000 -> 0.260776 | 0.939 / 1.062 | 0.500922 -> 0.455545 |

This improves the selected solver-native L1 diagnostic, but it does not improve the global maximum absolute coefficient or maximum L2 coefficient in the reviewed 10/19-step cases. Those limits are documented rather than hidden.

MiniMax-H3 still uses one audiovisual clock. Returned video sigma is one representation of a shared base-flow coordinate; ModelSamplingAV remains authoritative for the corresponding audio state. Runtime shift, audio_shift, and multiplier are read from the loaded model and are not hardcoded.

Inspect schedules without running the model with:

python tools/inspect_sa_schedule.py \
  --comfyui-path /path/to/ComfyUI \
  --comfyui-revision 8a33128f2f8c5585c57486c07de481241e70a39c \
  --steps 10 \
  --output /tmp/sa-schedule-10.json

The report includes returned/effective sigma, shared base time, implied audio sigma, lambda, interval deltas, interval ratios versus simple, node displacement, and native Adams diagnostics. It also keeps the media-failed lambda schedules available as diagnostic-only comparators so their extreme base-time distortion remains reproducible.

The node remains experimental. The bounded candidate has passed synthetic validation only; it must still pass matched decoded MiniMax-H3 media testing before any default or release recommendation changes. simple_control remains the default.

[Legacy/Research] MiniMax H3 RefDelta Scheduler

The custom profile-density scheduler is deprecated for ordinary production workflows. Existing saved workflows continue to work, and the scheduler/profile code remains available for explicit research.

For production, use ComfyUI:

BasicScheduler
scheduler = beta

The bundled r1024_provisional profile is a neutral beta-prior compatibility profile. There is no production calibration file to select from the FL2VA/Ref2VA replay workflow.

tools/build_profile.py is therefore research-only. It:

• strips comparison_* and ref_delta_* fields before scheduler binning;
• deduplicates replay copies of the same production telemetry;
• emits neutral difficulty by default;
• requires --experimental-stability-density for any non-neutral density;
• marks generated profiles as non-production research output.
• preserves version-1 beta-prior output by default and emits version-2 shared-base-time density only with --shared-flow-density.

Example research use:

python tools/build_profile.py \
  ComfyUI/output/refdelta_telemetry/*.csv \
  --output /tmp/r1024_scheduler_research.json \
  --id r1024_scheduler_research \
  --experimental-stability-density

Do not use Ref2VA distance as a scheduler error metric.

Risk/evidence model

RefDelta keeps two histories:

• solver history: every value returned to ER-SDE, including Spectrum forecasts;
• evidence history: actual model evaluations only.

Adaptive ER stage order uses deterministic trajectory risk. Stochastic adaptation uses combined trajectory/stochastic evidence. Stochastic pressure is measured from the native pre-gate ER-SDE increment relative to real denoised movement, preventing the controller from reducing its own future evidence.

Spectrum forecasts may consume the latest actual-only risk, stochastic pressure, and cached stability map, but they never become motion/diffusion/EMA/x0 evidence. Spectrum receives the exact final stochastic increment actually added to the state.

Telemetry and stability maps

When debug_telemetry is enabled, scalar JSONL/CSV telemetry is written under:

ComfyUI/output/refdelta_telemetry/

It includes trajectory/stochastic risk, native/applied stochastic values, controller mode, dynamic/static/audio targets, applied video-gate distribution, restore-mask distribution, motion/diffusion summaries, progress/source state, and slew activity.

When both debug_telemetry and debug_stability_maps are enabled, nonterminal actual steps also write compressed NPZ maps containing:

• temporal motion ratio;
• diffusion-change ratio when available;
• final restore mask;
• final applied video gate;
• seed, step, sigma, video shape, and controller config metadata.

Forecasts never emit evidence maps.

Spectrum compatibility

The reviewed baseline is ComfyUI-Spectrum-MiniMax-H3 v0.2.20+.

Important invariants:

• forecasts stay in solver history but out of RefDelta evidence history;
• the immediate forecast after an actual step consumes that actual step's latest stochastic evidence;
• forecast gaps preserve the last valid actual native stochastic/movement ratios;
• trajectory correction may target a forecast but uses only actual anchors/derivatives;
• spatiotemporal stability evidence updates only on actual model calls;
• Spectrum receives the exact post-controller stochastic increment;
• missing/stale/mismatched bridge state fails explicitly.

Static checkpoint context

The published rank-1024 sidecar shows a mixed approximation rather than a uniform rank limit: exact vector/bias pieces coexist with rank-limited matrix adapters. Overall retained energy is dominated by exact/small tensors, while actually compressed matrices retain much less of the original delta. This motivates trajectory-aware handling, but it does not make Ref2VA distance a quality metric or scheduler oracle.

Development

python -m pip install pytest ruff
python -m pytest -q
python -m ruff check .
python -m compileall -q comfyui_refdelta_solver tests tools

CI also checks native ER-SDE, SEEDS, SA-Solver, and scheduler fixtures across pinned ComfyUI revisions and builds the installable wheel.

License

GPL-3.0-or-later. The ER-SDE implementation is derived from GPL-licensed ComfyUI sampling code; see LICENSE.