MiniMax H3 Flow-Aligned Regenerate v0.2.1

MiniMax H3 Flow-Aligned Regenerate v0.2.1

v0.2.1 adds CI-gated Comfy Registry publication. Runtime and sampling behavior are unchanged from v0.2.0.

Distribution

• Adds a dedicated Publish to Comfy registry workflow using the pinned Comfy-Org/publish-node-action v1 revision already used by the companion MiniMax H3 latent-upscaler release path.
• Registry publication runs only after a successful CI push run on main and only when the tested commit changes the package version relative to its first parent.
• Manual workflow_dispatch remains available for explicit operator-controlled publication/retry.
• The workflow checks out the exact CI-tested commit, uses read-only repository permissions and non-persistent checkout credentials, and consumes REGISTRY_ACCESS_TOKEN only in the publish step.
• Publication is serialized per tested commit so overlapping workflow-run/manual attempts are not cancelled mid-publish.

Versioning

The package version is bumped to 0.2.1 rather than republishing 0.2.0 from a different source commit. This keeps the GitHub release/tag and Comfy Registry package version aligned to the same source revision.

Compatibility

No Python runtime, sampler, handoff, guidance, Continuum, Spectrum, metrics, or learned-upscaler integration code changes are included in v0.2.1.

MiniMax H3 Flow-Aligned Regenerate v0.2.0

MiniMax H3 Flow-Aligned Regenerate v0.2.0

v0.2.0 adds an optional learned 3D transfer at the Progressive Target Input handoff boundary, backed by the versioned API-v1 provider from Comfyui_Minimax_h3_latent_Upscaler.

Learned progressive handoff

• Adds handoff_transfer=learned_3d to MiniMax H3 Progressive Handoff (Target Input).
• Keeps handoff_transfer=bicubic as the compatibility default for existing workflows.
• Consumes the API-v1 H3_LATENT_UPSCALER clean-video provider rather than importing sibling-node internals.
• Replaces only the exact-probe clean-video spatial transfer. Deterministic target noise, conditional re-noising, caller audio, masks, target conditioning, sampler/Spectrum history boundaries, and the mandatory first high-grid actual H3 call remain unchanged.
• Learned CNN work is tracked separately and does not increase H3 model-call/NFE counters.

The coordinated provider is released in Comfyui_Minimax_h3_latent_Upscaler v0.2.0. The exact provider revision validated by this release is bdc670e5926bcefbe4022e17fe8b171fbfcf15de.

Decoded-media validation

The learned-transfer path was validated on aggressive progressive transitions around 1 MP rather than only on synthetic shape/contract tests.

• Around a 1152×864 (~0.995 MP) target, an aggressive bicubic handoff from roughly a ~0.55 MP private source showed substantial body/spatial handoff artifacts in the tested difficult prompt.
• Replacing only that boundary with learned_3d fixed the majority of the observed artifacts.
• source_scale=0.70 resolved to 800×608 → 1152×864 and was judged excellent.
• source_scale=0.65 resolved to 736×576 → 1152×864 and began losing reference likeness / tonal stability, so it is not promoted.
• The final higher-resolution gate used source_scale=0.70 at 832×640 → 1184×896 (~1.061 MP). The generated action differed, but decoded quality was again judged very good.
• That final run preserved 38 logical / 28 actual H3 NFE / 10 Spectrum forecast calls across two physical chunks, with 2 exact probes, 6 progressive sampler invocations, 4 history boundaries, copied audio, rebuilt high-grid conditioning, and an actual first high-grid H3 call in both chunks.
• BF16 CUDA learned inference took about 0.60 s and 0.77 s for the two chunks and added zero H3 NFEs.

For this prompt around 1 MP, source_scale=0.70 is the current tested quality/compute point and 0.65 is below the observed fidelity floor. These values are not claimed as universal optima.

Guidance interpretation

The latest learned-transfer media sweep used direction+acceleration. It does not establish an acceleration advantage over direction-only. The earlier matched evidence still supports direction as the conservative preferred guidance mode; acceleration, temporal correspondence, downsample consistency, auto handoff, and resolution-aware refine sigmas remain experimental controls unless future matched media shows a clear benefit.

Compatibility and validation

• Existing bicubic Progressive Target Input workflows retain their prior behavior.
• Full 1-to-0 H3 sigma schedules remain required for progressive handoff.
• CI validates Python 3.10–3.13, Ruff and formatting, 142 unit/synthetic tests, compileall, package build, isolated wheel import, and pinned native/sibling source contracts.
• The source-contract gate pins the merged learned-provider revision and checks its API/kind constants, helper, provider classes, callable surface, and ComfyUI type.

v0.1.0

Initial public release of MiniMax H3 Flow-Aligned Regenerate.

Highlights

• H3-native low-resolution trajectory capture and time-aligned high-resolution guidance.
• Progressive Target Input handoff for Continuum: early H3 sampling can run on a private lower-resolution grid before switching to the final grid without a separate learned-refine replay.
• Integrated Continuum refine-state guidance for the existing MiniMax H3 Latent Upscaler + Refine path.
• Experimental resolution-aware refine sigmas, temporal correspondence, acceleration, downsample consistency, reference-budget diagnostics, and attention diagnostics.
• Runtime metrics distinguish logical sampler calls, actual H3 NFEs, Spectrum forecasts, handoff probes, guidance events, and resolution-map events.
• CI covers Python 3.10-3.13, source-contract checks, Ruff, formatting, 133 tests, compilation, package build, and isolated wheel validation.

Tested operating point

The strongest matched difficult-motion result tested during development used:

• 14 SA-Solver-PECE outer steps
• 736x736 private source -> 896x896 target
• source_mode=scale, source_scale=0.83
• fixed handoff 0.35 -> actual ~0.358 / index 9
• direction guidance 0.25
• acceleration / temporal / consistency 0
• 54 logical calls / 36 actual H3 NFEs / 18 Spectrum forecasts across two chunks
• effectively 9 low-grid + 5 high-grid outer steps per chunk

D12 was judged better than D10, and D14 slightly better again. These are tested quality/speed tradeoffs, not universal optima.

Decoded-media validation

Completed smoke coverage includes:

• integrated two-pass C7=7+7, C6=7+6, C5=7+5, plus exploratory C4=7+4;
• Progressive Target Input at D10/D12/D14;
• fixed and auto-computed handoff selection;
• corrected HiFlow-style acceleration;
• conservative temporal correspondence;
• downsample consistency;
• matched E0/E1 resolution-aware learned-refine sigma mapping with the real learned refiner enabled in both arms.

Acceleration, downsample consistency, auto handoff, D14 temporal guidance, and resolution-aware refine sigmas were structurally valid but did not show a clear decoded-media advantage in the matched difficult-motion tests, so they remain non-promoted experimental controls.

Earlier temporal runs decoded with the accidental TensorRT w4a16_awq VAE are excluded from temporal-quality attribution.

Scope

This project is a training-free research implementation informed by public work. It does not reproduce MiniMax's closed H3-Regenerate-2K model or unreleased sparse-attention topology. Broad cross-prompt quality claims are intentionally withheld; the repository documents the larger optional benchmark matrix for publication-grade evaluation.