MiniMax-H3-Flow-Aligned-Regenerate Introduction
MiniMax-H3-Flow-Aligned-Regenerate is a collection of ComfyUI nodes that helps you reuse a lower-resolution MiniMax H3 generation when creating a higher-resolution result. It also provides dedicated tools for continuing a video across multiple chunks while preserving the accepted audio and video overlap.
Instead of asking H3 to do every stage of generation at the final resolution, you can establish the scene on a smaller working grid, transfer that work to a larger grid, and refine it there.
For you as an AI artist, this provides three main possibilities:
- Reuse an established composition: guide a higher-resolution pass using information from the lower-resolution generation, rather than relying only on its final frame or final latent.
- Move between resolutions during generation: begin on a smaller grid and switch to your target resolution within one sampling schedule.
- Continue longer scenes: use the recommended Native Source Continuation path with H3 Continuum-Plus to carry forward source information and protect the accepted audio/video overlap.
The extension is training-free: you do not need to train a new model to use its guidance or handoff nodes. However, the recommended learned-transfer workflow uses a separately trained latent upscaler.
What it can—and cannot—do
The extension is designed to improve how generation work is reused across resolutions and continuation chunks. It does not guarantee identical composition, seamless joins, perfect camera behavior, or a particular speed improvement for every prompt.
Higher-resolution H3 refinement still needs enough GPU memory for the target-resolution stage. Starting at a smaller resolution does not eliminate that requirement.
The author describes this project as an independent research implementation informed by public work. It does not reproduce MiniMax’s closed H3-Regenerate-2K implementation or an unreleased sparse-attention model. No Keyless implementation is included.
How MiniMax-H3-Flow-Aligned-Regenerate Works
A few useful terms
| Term | Plain-language meaning |
|---|---|
| Latent | A compressed representation of your video that the model works on before it is decoded into visible frames. |
| Sampling / denoising | The sequence of model steps that gradually turns noise into generated content. |
| Trajectory | The intermediate predictions made during that sequence—not just the final result. |
| Source grid | The smaller working resolution used during the early stage. |
| Target grid | The final working resolution. |
| Protected prefix | The beginning of a continuation chunk copied from the preceding chunk and marked to remain unchanged. |
| VAE decode | The step that converts video latents into viewable frames. |
Approach 1: flow-aligned two-pass guidance
Think of a low-resolution generation as a sequence of rough sketches. The last sketch is useful, but the earlier sketches also show how the scene developed.
The two-pass path records those intermediate H3 predictions and uses matching states to guide a later higher-resolution pass.
Low-resolution H3 generation
|
Trajectory Capture
|
Captured trajectory ---------------------+
|
Learned upscale / refine initialization --+
|
Flow-Aligned Regenerate
|
Higher-resolution H3
A typical workflow uses:
- MiniMax H3 Flow Trajectory to hold the shared trajectory.
- MiniMax H3 Trajectory Capture during the first pass.
- A learned latent upscale/refine initialization.
- MiniMax H3 Flow-Aligned Regenerate during the second pass.
For a Continuum refinement workflow using refine_state, use MiniMax H3 Flow-Aligned Refine State.
Recommended starting guidance for this explicit two-pass approach: direction.
Approach 2: progressive handoff
Progressive handoff works more like drawing a composition on a small canvas, enlarging the working canvas partway through, and finishing at the larger size.
The normal progressive path:
- Performs early sampling on the reduced source grid.
- Takes an exact handoff probe.
- Transfers the predicted-clean video to the target grid.
- Continues H3 refinement at the target resolution.
The learned spatial transfer affects video only. Audio does not pass through the latent upscaler.
Recommended continuation: reuse what was actually rendered
MiniMax H3 Native Source Continuation is the recommended progressive interface for both single chunks and Continuum sequences.
For continuation, it reuses the preceding chunk’s already-computed source clean prediction, paired with the exact returned target suffix. Its default target-band reconciliation adjusts the broad source layout to agree with the video that was actually returned.
An analogy is keeping both a rough sketch and a finished painting: before continuing the sketch, you align its major shapes with the finished painting so the next section follows the scene viewers actually saw.
The learned upscaler then transfers the generated trajectory to the target grid. High-resolution refinement uses the original protected target prefix, and the returned protected audio/video prefix remains authoritative.
Important distinctions:
- Native source carry adds no extra H3, upscaler-provider, or VAE calls.
- A learned handoff still occurs as part of the progressive workflow.
- Missing or mismatched native source context stops generation before low sampling; it does not silently switch to another continuation method.
- Exact overlap preservation does not guarantee that newly generated content immediately after the overlap will be visually seamless.
MiniMax-H3-Flow-Aligned-Regenerate Features
Main generation nodes
| Node | What it does | When to use it |
|---|---|---|
| MiniMax H3 Flow Trajectory | Holds the shared trajectory used by capture, guidance, and progressive execution. | When building trajectory-based workflows. |
| MiniMax H3 Trajectory Capture | Records first-pass predicted-clean H3 states. | On the low-resolution pass of a two-pass workflow. |
| MiniMax H3 Flow-Aligned Regenerate | Guides a later H3 pass using the matching captured state. | For explicit upscale-and-refine workflows. |
| MiniMax H3 Flow-Aligned Refine State | Provides flow-aligned guidance through Continuum’s refine_state. | When refining Continuum chunk outputs. |
| MiniMax H3 Progressive Handoff | Provides a generic source-sized resolution handoff. | For workflows defined from the smaller source input. |
| MiniMax H3 Progressive Handoff (Target Input) | Provides progressive control for workflows defined at the final target geometry. | For general target-input workflows; protected prefixes have a special fallback described below. |
| MiniMax H3 Native Source Continuation | Combines a uniform source trajectory, native source carry, learned transfer, and exact protected audio/video ownership. | Recommended for new single-chunk and Continuum progressive workflows. |
| MiniMax H3 Partitioned Exact-Prefix Handoff [Legacy Controls] | Preserves the older control layout and explicit continuation settings. | For existing saved workflows rather than new projects. |
| MiniMax H3 Continuum Decode Context | Supplies future latent context to the native temporal decoder at exact chunk joins. | To address the separate decoder-window boundary problem. |
Source-prefix projection choices
These settings determine how the reduced-grid prefix is prepared. None changes the authoritative target prefix.
source_prefix_projection | Behavior |
|---|---|
native_source_carry_target_band | Recommended new-node default. Uses broad spatial information from the resized returned target prefix and retains finer carried detail only where the latent structures agree. |
native_source_carry | Previous behavior: uses the preceding chunk’s reduced-grid estimate at the handoff noise level without target-band reconciliation. |
latent_bicubic | Reconstructs only the source prefix when paired native context is unsuitable. |
The author reports that target-band reconciliation removed a visible boundary and a background redraw shortly after the join in user-reviewed continuation renders. This is evidence from those renders, not a universal seam-quality guarantee.
Detail transport
uniform_source_detail_transport=true enables a checked detail correction before high refinement.
It uses the last protected prefix to estimate detail lost during learned transfer. The correction is applied only when its calibration and holdout checks pass; otherwise, the extension reports that it was skipped.
- It adds no H3, provider, or VAE evaluation.
- Its calibration still uses compute and memory.
- It has no prefix to correct in an entirely generated single chunk.
- Existing saved
falsevalues and legacy-node defaults remain unchanged.
Guidance modes and customization
| Control | Meaning and practical use |
|---|---|
direction | Conservative recommendation for explicit two-pass guidance. |
direction+temporal | Recommended progressive default, combining directional and temporal guidance. |
direction+acceleration | Research control that activates acceleration_weight. |
downsample_consistency | Research control that activates consistency_weight. |
direction_weight | Adjusts the contribution of directional guidance. |
temporal_weight | Adjusts the contribution of temporal guidance. |
low_frequency_cutoff | Controls the low-frequency band used by guidance. See the usage documentation for its exact interpretation. |
For example, changing acceleration_weight while using direction+temporal has no effect: that weight is inactive in this mode. Switch modes only when you intentionally want to test a different guidance method.
When comparing settings, keep your seed, prompt, references, model, resolution, duration, and sampler unchanged. Judge the decoded video as well as any metrics.
Target Input’s exact-prefix fallback
Progressive Handoff (Target Input) behaves differently depending on the mask:
- Unprotected or fractional-mask input: runs the normal progressive low/probe/high path.
- Exact protected video prefix: forwards the original target-grid noise, latent, mask, and schedule through one ordinary target-grid sampler lifetime.
In the exact-prefix fallback, there is:
- No private low-grid sampler.
- No exact handoff probe.
- No learned-upscaler call.
- No resolution boundary or progressive-history boundary.
For canonical exact-prefix continuation, Target Input also uses a validated four-tick audio-overlap ramp during sampling. The original exact video/audio mask remains authoritative at output.
This is separate from Native Source Continuation’s default sampler_mask_exact_timestep audio-overlap mode. Do not assume the two interfaces use identical continuation behavior.
Decoder context and diagnostics
Continuum Decode Context is placed immediately before normal Video VAE Decode. It supplies real future latent context at chunk joins without changing accepted sampling latents, continuation state, masks, audio, or the assembly plan.
For deeper investigation:
- MiniMax H3 Local Boundary Audit replays saved boundary bundles through the production native video VAE.
- MiniMax H3 Metrics JSON writes structured model, sampler, geometry, and forecasting metrics.
- Full-stage continuity replay saves matched PNGs and stage records without running H3 sampling.
Audit measurements help locate changes between stages. They do not automatically prove that a camera movement, focus change, or soft background is a defect.
Legacy and experimental controls
The following remain available for compatibility or controlled research:
| Node | Status |
|---|---|
| Progressive Mixed-Grid Continuum | Deprecated compatibility path; not recommended for new production work. |
| Progressive Target-Sparse Continuum | Experimental and not recommended for final output. Testing showed cascading skin, clothing, and background defects. |
| Refine Target Geometry | Experimental target-sizing metadata control. |
| Resolution-Aware Sigmas | Experimental resolution-dependent refinement schedules; off by default. |
| Reference Budget | Experimental reference diagnostics and guarded reference cap. |
| Attention Lab | Experimental attention and retention diagnostics. |
Retired live replay and kernel-observer controls may still occupy saved widget slots so older graphs load, but they no longer add sampling calls or hooks.
Using MiniMax-H3-Flow-Aligned-Regenerate
Choose an example workflow
Start from the loadable example workflows rather than rebuilding the graph from scratch.
| Example | Purpose |
|---|---|
| Native Source Continuation | Recommended single-chunk and Continuum interface with learned 3D transfer and exact protected audio/video. |
| Progressive Target Input | General progressive control defined at target resolution. |
| Progressive Source Input | Progressive control defined from the source input. |
| Partitioned Exact Prefix | Preserved legacy control layout. |
| Full-stage continuity | Diagnostic example for investigating continuity across stages. |
Matching .api.json graphs are included. Wiring overlays are specifications, not loadable workflows.
Recommended Native Source Continuation setup
- Use MiniMax H3 Native Source Continuation (
H3NativeSourceContinuationHandoff). - Connect VDN-H3-Plus partitioned transport and the learned 3D upscaler provider.
- Keep VDN’s full trained linear complement enabled.
- Connect the patched MODEL to both the scheduler and guider.
- For Continuum Native Masked generation, use that same patched MODEL for the initial chunk and every continuation.
A first chunk generated with a different MODEL cannot supply the required native source carry.
Native carry requires:
- Batch-one H3 video.
- Native
5k+2latent-token lengths. - An even spatial grid.
- A target grid larger than the reduced source grid.
These are latent-layout requirements, not a rule to apply directly to your visible frame count.
Recommended new-node defaults
| Setting | Default |
|---|---|
source_mode | scale |
source_scale | 0.70 |
source_width, source_height | 864, 640, used only in pixels mode |
handoff_coordinate | 0.35 |
handoff_selection | fixed |
guidance_mode | direction+temporal |
direction_weight | 0.25 |
acceleration_weight | 0.25, inactive in the default guidance mode |
consistency_weight | 0.25, inactive in the default guidance mode |
low_frequency_cutoff | 0.25 |
temporal_weight | 0.20 |
audio_guided_overlap_mode | sampler_mask_exact_timestep |
audio_guided_overlap_ticks | 16, with zero effective release in exact mode |
uniform_source_detail_transport | true |
source_prefix_projection | native_source_carry_target_band |
capture_boundary_witness | false |
source_scale=0.70 requests a reduced working size relative to the target. Switching to pixels mode makes the explicit source width and height relevant instead.
The shipped learned examples use:
minimax_h3_latent_upscaler_3d_bf16.safetensors- CUDA and
bf16 offload_after_upscale=falseres_multistep- No Turbo LoRA
Saved workflows retain their stored values. Updating the extension does not automatically apply these defaults.
Optional patch order
For the native-source Continuum stack, preserve this order when the optional patches are present:
DiffAid
-> Untwisting RoPE
-> Spectrum
-> Native Source Continuation
-> Continuum
DiffAid, Untwisting RoPE, and Spectrum are optional. Sol-H3 is also optional; the continuation path supports the selected ComfyUI attention backend through VDN-H3-Plus.
Spectrum is a separate approximate accelerator. It forecasts some H3 feature states instead of running every full transformer evaluation, so it can change the output even with the same seed. It is not the same mechanism as Flow’s resolution handoff.
MiniMax-H3-Flow-Aligned-Regenerate Models
Does the extension include its own model variants?
The extension provides workflow nodes, not a separate family of generative checkpoints. Your MiniMax H3 model generates the content; the recommended companion upscaler enlarges its video latents.
Learned 3D upscaler checkpoints
The companion MiniMax H3 Latent Upscaler-Plus uses checkpoints available from LBH-123-AI on Hugging Face.
| Checkpoint | Stored precision | Relevance |
|---|---|---|
minimax_h3_latent_upscaler_3d_bf16.safetensors | BF16 | Used by Flow’s supplied learned-transfer examples and preselected by the provider when available. |
minimax_h3_latent_upscaler_3d_fp16.safetensors | FP16 | An alternative stored-precision version of the same 3D architecture. |
minimax_h3_latent_upscaler_3d_fp32.pth | FP32 | Full-precision storage version of the same architecture. |
These are not different artistic styles. The provider’s precision setting controls inference precision independently of the checkpoint’s stored precision.
The companion package also has a 2D upscaler node, but Flow’s recommended learned handoff uses the 3D provider. A 3D checkpoint is not compatible with the 2D node.
Learned transfer versus bicubic
The general unprotected Target Input handoff supports:
learned_3d: transfers the clean video estimate using the trained upscaler; used in the supplied target-input workflow.bicubic: an interpolation-based alternative.
This transfer selection is distinct from the latent_bicubic source-prefix projection option.
The learned upscaler preserves the time axis and enlarges only the spatial dimensions. It never processes audio.
What’s New with MiniMax-H3-Flow-Aligned-Regenerate
v0.3.14 — coordinated release, 17 October 2026
The recommended Native Source Continuation workflow remains unchanged.
| Change | Why it matters |
|---|---|
| PR #107 retains Sol-H3 grouped K/V fusion in continuation. | Keeps the coordinated Sol/VDN continuation integration available without replacing VDN’s required trained linear work. |
| PR #102, #103, and #108 expand diagnostics. | Provide full-stage tone evidence, target-sized comparisons, and per-prompt lifecycle/paging records for investigating continuity and memory behavior. |
| PR #105 fixes old local-audit ownership retention across prompts. | Prevents completed audit state from unnecessarily retaining model owners across later prompts. |
Lifecycle tracing is optional. Set H3_LIFECYCLE_TRACE=1 before ComfyUI starts when you need those records.
Coordinated companion versions
| Component | Release paired with Flow v0.3.14 |
|---|---|
| Flow-Aligned Regenerate | v0.3.14 |
| Sol-H3 | v0.1.11, optional |
| VDN-H3-Plus | v1.5.10 |
| H3 Continuum-Plus | v3.4.9 |
Later supplied Continuum notes identify v3.4.10 and v3.4.11 as compatible with Flow v0.3.14. Those releases address speech-conditioning isolation and package publishing identity; they are not Flow releases.
Earlier documented changes
- v0.3.13: includes PR #101. The supplied summary does not describe its individual changes; consult the release history.
- v0.3.11: established
progressive_uniform_sourceas the default partitioned-continuation profile and supports continuation with the selected ComfyUI attention backend rather than requiring Sol-H3.
For the complete historical record, see Release Notes and Release Context. Experimental sibling-package features are not automatically part of this release.
Troubleshooting MiniMax-H3-Flow-Aligned-Regenerate
Continuation stops because native source context is missing or mismatched
- Confirm that the initial chunk was generated using the same Native Source Continuation-patched MODEL.
- Check that the same patched MODEL reaches the scheduler, guider, and Continuum sequence.
- Verify the required batch, temporal layout, even spatial grid, and source/target size relationship.
- Regenerate the initial chunk with the correct graph if it came from a different MODEL.
Stopping is intentional: native carry does not silently substitute an unrelated source estimate.
A join shows background redraw or a visible boundary
- Check
source_prefix_projectionin your saved node. - If it uses the previous
native_source_carrybehavior, selectnative_source_carry_target_band. - Check whether detail transport is enabled and whether its correction passed or was skipped.
- Compare the decoded video around the join—not only the overlap itself.
- If the problem remains, use full-stage continuity replay or a boundary audit to identify the stage where the change appears.
Target-band reconciliation does not guarantee every join. Camera, focus, prompt, and reference behavior remain model-dependent.
The continuation becomes slightly darker
Slight continuation darkening remains a known issue.
- Compare matched frames before and after the join.
- Use full-stage diagnostics to distinguish sampling-stage changes from decoder or assembly behavior.
- If you use Continuum-Plus, review its Auto 3 (Sustained Tone) video-seam option. It can reduce sustained darkening in suitable scenes, but it modifies newly generated decoded frames and is not a geometric repair.
- For a controlled comparison, keep generation settings fixed and compare seam modes such as
AutoorAnalyze Only.
The upscaler is connected but does not run
- Identify the handoff node.
- If you use Target Input with an exact protected video prefix, zero upscaler calls are expected: this path uses the target-grid fallback.
- For an unprotected learned handoff, check that
learned_3dis selected and the 3D provider is connected. - For the recommended native-source continuation workflow, use the dedicated Native Source Continuation example.
Changing a setting has no visible effect
- Confirm that the setting is active in the selected mode.
- Remember that
source_widthandsource_heightapply only inpixelsmode. acceleration_weightis active only indirection+acceleration.consistency_weightis active only indownsample_consistency.- Compare saved widget values with current defaults; existing workflows are not rewritten on update.
Retired replay flags cannot add sampling calls or hooks, even if their saved slots remain present.
GPU memory runs out during high-resolution refinement
- Remember that the high stage still runs H3 at the full target resolution.
- Reduce target resolution or duration and test again.
- Review the companion upscaler’s
offload_after_upscaleoption if keeping its network resident contributes to memory pressure. Offloading adds transfers and can reduce speed. - Keep VDN’s trained linear complement enabled; disabling required model work is not a supported memory shortcut.
- For a local audit, use the same production VAE and settings. The audit stops when decode memory is insufficient and does not retry a managed decode after OOM.
No universal minimum VRAM requirement or speedup is established by the supplied documentation.
A decoder-window join remains visible
- Distinguish the decoder boundary from a change already present in sampling latents.
- Place MiniMax H3 Continuum Decode Context immediately before normal Video VAE Decode.
- Follow the Decode Context guide.
This helper addresses a separate decoder-window issue. It does not replace sampler-time prefix protection or audio-overlap behavior.
A Local Boundary Audit cannot find its bundle
- Point
bundle_pathto the exported directory containingmanifest.jsonand the full binary operands, or selectmanifest.jsonitself. - Connect the same native video VAE instance and settings used for production.
- Set
chunk_join_frameto the assembled output’s first new frame, or leave it at zero for labels relative to the join. - In WSL, use a Linux path such as
/home/..., or a matching\\wsl.localhost\<distribution>\.../\\wsl$\<distribution>\...path.
Windows Explorer paths are translated only when the distribution matches WSL_DISTRO_NAME. Quotes added by Copy as path are accepted.
Frequently asked questions
Do you need Sol-H3?
No. Sol-H3 is optional. Native Source Continuation requires VDN-H3-Plus partitioned transport and the recommended learned upscaler, but supports the selected ComfyUI attention backend.
Why does Sol-H3 show no custom-kernel acceleration on native Windows?
The supplied Sol documentation states that its custom SOL and Exact Runtime kernels do not execute on native Windows. Those nodes delegate to inherited/native paths. Supported custom-kernel targets are Linux/WSL2 on SM120/SM121 hardware.
Does exact prefix preservation mean the whole continuation is identical?
No. It preserves the accepted protected overlap. Newly generated frames can still change tone, background detail, motion, or framing.
Does this upscale or regenerate audio?
The spatial upscaler processes only video. Audio behavior is managed by the sampling and continuation contracts.
Does a successful detail-transport calibration guarantee a seamless result?
No. It permits that correction to run; final visual acceptance still requires watching the decoded output.
Can you keep using an older graph?
Legacy node identifiers and saved layouts are retained where documented. For new work, use Native Source Continuation rather than deprecated Mixed-Grid or experimental Target-Sparse paths.
Will the same settings always produce the same performance?
No. Timing depends on geometry, duration, sampler, model residency, references, Spectrum policy, hardware, and optional attention backends.
Learn More about MiniMax-H3-Flow-Aligned-Regenerate
Start here
- Usage guide — wiring and parameter details.
- Workflow wiring — how the example graphs connect.
- Example workflows — loadable starting points.
- Native source-prefix carry — continuation requirements and behavior.
- Target-band reconciliation — how the default source-prefix projection works.
Understanding joins and output quality
- Continuum Decode Context — decoder context at chunk boundaries.
- Full-stage continuity replay — matched stage comparisons and tone investigation.
- Boundary audit replay — replaying saved evidence with the native VAE.
- Decoded-media validation — recorded output-validation evidence.
- Performance evidence — workflow-specific timing results.
Companion tools
- MiniMax H3 Latent Upscaler-Plus — learned video-latent enlargement and the 3D handoff provider.
- VDN-H3-Plus — trained hybrid-attention integration and partitioned transport.
- H3 Continuum-Plus — chunked generation, Native Masked continuation, references, audio routing, and assembly.
- Spectrum MiniMax H3 — optional approximate feature-forecasting acceleration.
- Sol-H3 — optional attention/runtime integration with hardware-specific requirements.
Advanced and historical documentation
- Legacy continuation controls
- Deprecated Mixed-Grid documentation
- Historical Mixed-Grid seam investigation
- Target-Sparse research path
- Architecture
- Research rationale
- Credits and provenance
- Release history
MiniMax-H3-Flow-Aligned-Regenerate License
MiniMax-H3-Flow-Aligned-Regenerate is licensed under the Apache License 2.0.
Copyright 2026 xmarre.
Companion repositories, referenced papers, and third-party model weights retain their own copyrights and licenses.
