0.6.63 — Cumulative Director and RefMod update

ComfyUI-MiniMax-H3-LongMedia 0.6.63 — cumulative Director and RefMod update

Description

This cumulative release brings together the changes from 0.6.60 through 0.6.63: Director timeline and regeneration controls, RefMod reliability improvements, more consistent MultiClip refinement, workflow-persisted presets, and preview-memory fixes.

Director and timeline

• Enable or disable individual generated MAIN clips for RUN. Disabled clips keep their current TAKE; partial runs require compatible cached MultiClip state.
• Merge adjacent approved clips into a reversible timeline block while retaining their individual TAKEs and per-clip RUN settings.
• Improve timeline gap creation, trimming, source-time display, clip drops, cut markers, and selective regeneration. CUT markers remain separate from Knife splits.
• Sort TAKEs and preserve the gallery's scroll position while selecting a take.
• Save shared Camera presets from any CAMERA clip, then apply, overwrite, or delete them from the shared library. Presets persist in director_json and are available to existing and newly created CAMERA clips.

RefMod and MultiClip

• Create RefMods from up to 128 image, video, or audio sources, with per-source selection and crop, clipboard image paste, and token estimates. Extraction controls now reach the relevant encoder stages.
• Add concept-specific RefMod guidance and workflow-saved scene-direction presets. Character RefMods retain identity authority across MultiClip boundaries.
• Route long interval-gated RefMod attention through streamed-Q INT8 when needed, retaining globally quantized K/V and per-query masks without a dense Q×K allocation. Full-span identity masks can keep the selected attention backend.
• Reuse original uploaded media for RefMod ENCODE; existing .safetensors artifacts without source files remain usable.
• Carry refined high-resolution overlap between MultiClip segments to reduce geometry and color discontinuities.

Runtime, presets, and media

• Save Sampler presets in the workflow, report when settings differ from the selected preset, and keep Seed independent. motion_repair also survives workflow reloads.
• Release detached preview decoders, load TAKE thumbnails lazily, and bound decoded preview-image retention. Timeline Fit and layer-height controls are easier to reach.
• Keep BASE and VIDEO reference scales with the clips that own those references, and bypass H3 sampler preparation for timelines made only of immutable MEDIA clips.
• Preserve the original audio assembly path and exact audio passthrough during MultiClip refinement.

0.6.63 fixes

• Treat absent optional ComfyUI conditioning branches as empty lists before guider preprocessing. Missing required positive conditioning now raises an explicit error instead of a NoneType iteration traceback.
• Changing an AUDIO clip role from diegetic to reactive remains supported and may invalidate that clip's TAKE for regeneration.
• Add a shared, editable Camera preset library to the Director document.

Validation

Python syntax for 45 files, JavaScript syntax for seven files, three packaged JSON files, ZIP integrity, and installed/source file hashes were checked. The full test suite and a live ComfyUI/GPU render were not run.

Documentation

• Russian release notes
• English release notes
• The package includes the bilingual Director guides and parameter references.

Asset

Attach ComfyUI-MiniMax-H3-LongMedia-0.6.63.zip.

v0.6.60 — RefMod, Director and MultiClip

ComfyUI-MiniMax-H3-LongMedia 0.6.60 — RefMod, Director and MultiClip fixes

Description

LongMedia 0.6.60 brings the latest Director improvements and H3 runtime fixes together, with a complete Russian and English parameter reference based on the fields in this build. The package version remains 0.6.60.

Highlights

• Added complete parameter references for Setup, Director, RefMods, Planner/Cameras, Video Reconstructor, Sampler and Decode. The guides distinguish visible controls, mode-dependent fields, hidden compatibility values and legacy inputs that are ignored.
• Added a RefMod Sources builder for up to 128 image, video and audio sources, including clipboard image paste and per-source/total visual-token estimates.
• RefMod extraction controls now describe and reach the correct stages: short-edge downscale before VideoVAE, visual-token cap, aspect-fitted compressed grid, video frame limit, audio-only voice duration and Compressed-only latent refinement.
• Fixed RefMod attention routing so a full-span exact-strength reference can preserve the selected backend, including zero-copy SLA, when the merged mask is mathematically an identity operation. The Sampler also preserves the selected ComfyUI SLA hook when cloning guider options.
• Character RefMods retain identity authority across MultiClip boundaries without altering audio or other RefMod routes.
• MultiClip Latent Hi-Res + Refine now carries refined overlap context from one generated clip into the next Stage 2 pass while preserving Stage 1 continuation and exact audio.
• Director timeline fixes include true-gap-only gap controls, gap duration display, Fit after horizontal scrolling, clip-owned Cut markers, separate Cut and Knife behavior, trim source-time hints, TAKE sort controls and gallery position preservation.
• Improved Director preview lifecycle by releasing detached filmstrip decoders, unloading old TAKE thumbnails, lazy-loading gallery previews and bounding decoded preview images with an LRU cache.
• Fixed motion_repair workflow persistence, immutable MEDIA-only rendering and several RefMod/Director import, routing and state issues.

Important behavior

• RefMod Concept Type adds broad prompt guidance. Its Description is library metadata; neither parameter guarantees style transfer or changes the extracted latent by itself.
• RefMod ENCODE prepares native H3 VAE latents. It does not run the DiT or Sampler.
• RefMod Refinement steps apply only to COMPRESSED extraction. Sampler Stage 2 uses the separate Refine Sigmas input; legacy refine_steps is ignored.
• Director RefMods are active only during their timeline intervals. Stretch the layer clip across the timeline for a full-scene reference.
• Start with sampler_mode=auto, memory_mode=auto and attention_mode=auto for normal workflows.

Documentation

• Russian: docs/README_RU.md
• English: docs/README_EN.md
• Full parameter reference: docs/PARAMETER_REFERENCE_RU.md and docs/PARAMETER_REFERENCE_EN.md
• Release notes: docs/RELEASE_NOTES_0.6.60_RU.md and docs/RELEASE_NOTES_0.6.60_EN.md

Asset

Attach the ZIP whose filename starts with ComfyUI-MiniMax-H3-LongMedia-0.6.60-.

v0.6.54 — Native Continuation, Selective Regeneration & Runtime Stability

ComfyUI-MiniMax-H3-LongMedia v0.6.54

This release consolidates the validated work developed after the public v0.6.50 baseline. The main theme is simple: already-approved timeline media should stay approved.

Native continuation and immutable MEDIA

Director now distinguishes H3 GENERATED clips from immutable imported MEDIA.

GENERATED -> GENERATED reuses the saved native TAKE continuation state and samples only the new child clip.

MEDIA -> GENERATED VAE-encodes only the required external tail and attaches it as a native frame-0 H3 keyframe to a fresh target.

Imported video never becomes a diffusion render unit, never gets copied into target x0, and never takes a full-video VAE roundtrip just to extend the timeline.

Final assembly decodes GENERATED runs only and inserts original MEDIA RGB/audio directly.

Selective regeneration

Regenerate Clip keeps compatible cached prefix/suffix TAKEs and samples only the selected GENERATED clip.

Regenerate From Here keeps the validated prefix and samples only the dependent GENERATED suffix.

Appending a new clip reuses the previous approved TAKE instead of rebuilding the timeline from clip 1.

Cached clips skip redundant text/reference conditioning during selective Setup preparation.

Audio continuation and mixed timelines

Per-BASE audio continuation supports AUTO, CONTINUE and FRESH. External MEDIA contributes only the required audio tail as context; its original soundtrack remains untouched in the old timeline section. Mixed mono/stereo assembly now uses an explicit [1,C,L] contract with exact mono-to-stereo duplication when needed.

Quantized runtime / Windows

INT8/W4A8/NVFP4 AUTO memory policy now uses physical packed storage + activation headroom instead of BF16-equivalent logical model size.

Explicit manual sampler values remain authoritative in NORMAL/AUTO-normal.

Large Windows safetensors remain file-backed through ComfyUI ModelMMAP / TensorFileSlice whenever available, preserving DynamicVRAM/AIMDO and avoiding eager whole-model RAM loading.

Setup releases TE/transient conditioning state before diffusion sampling without aggressive working-set/file-cache trimming.

Compatibility

The existing Director NLE/TAKE workflow, camera/embedding timing, Refiner AV contract and H3 quality path are preserved. Stage 2 remains video-only with exact Stage-1 audio passthrough. SPEED/progressive-resolution is not included in this release.

Full details: docs/RELEASE_NOTES_0.6.54_EN.md / docs/RELEASE_NOTES_0.6.54_RU.md.

v0.6.50 — Director Timeline Editing & TAKE Workflow

ComfyUI-MiniMax-H3-LongMedia v0.6.50 — Director Timeline Editing & TAKE Workflow

This release focuses on the Director editing experience and one Stage-2 Refiner device-contract fix. The notes below contain only changes introduced after the public GitHub 0.6.42 release.

Director becomes a real timeline editor

• TAKE management now supports checkbox multi-select, Select All/Clear, batch delete, folder-tree navigation and MOVE TO for moving one or many TAKEs into existing folders.
• Layers are first-class selectable editor objects with rename, duplicate/delete where valid, enable/lock/mute controls and layer properties.
• Every track can be resized independently by dragging its separator; track heights persist in Director layout state.
• Timeline/Inspector/Global Prompt layout reflows correctly after resizing, removing the large vertical gaps that could appear between panels.
• The toolbar is cleaner: icon groups remain visually separated, duplicate import actions are removed, and Selected TAKE now sits directly above the timeline.

NLE-style timeline editing

• Added custom timeline context menus with Duplicate, Copy, Cut, Paste, Delete, Knife clip here and Knife all layers here, plus matching keyboard actions.
• Knife clip here now cuts only the clicked clip or block. The global knife remains a separate explicit operation.
• Clips and blocks can be positioned more freely. Magnet snapping remains available, with reduced snap tolerance and Alt-drag to temporarily bypass snapping.
• BASE clips can be trimmed or moved to create real timeline gaps instead of being forced back into a contiguous chain.
• Right-clicking an empty interval exposes a type-aware Fill Gap action for BASE, CAMERA, AUDIO and custom layers.
• A blank creation zone below the tracks provides quick creation of Prompt, Embedding, Character, Reference, Video and Audio layers plus applicable core timeline blocks.
• Global scissors split every unlocked block crossing the playhead, while MultiClip + CLIP stays MAIN-only.

Refiner reliability

• Fixed mixed CPU/CUDA AV noise ownership on the high-VRAM/native path. Frozen Stage-1 audio noise now follows ComfyUI's stock prepare_noise() device/dtype contract in both exact and windowed Stage-2 refinement, preventing nested-noise device mismatch before sampling starts.

Compatibility

The release keeps the existing 0.6.42 high-VRAM/native H3 compute behavior and Sampler controls unchanged; those changes are not repeated here because they are already part of the public 0.6.42 release.

v0.6.42 — High-VRAM Performance & Native H3 Compute Path

ComfyUI-MiniMax-H3-LongMedia v0.6.42

This release focuses on high-VRAM GPU performance and fixes Sampler controls that could previously be silently overridden by LongMedia's internal memory policy.

High-VRAM performance fixes

• mlp_chunk_tokens is now respected in normal memory mode.
• Large manual values such as 16384, 32768, 65536, and 131072 are no longer clamped to 8192.
• mlp_chunk_tokens = 0 now genuinely disables LongMedia MLP chunking.
• Manual VRAM guard values set to 0 now remain disabled instead of being replaced by internal minimum values.
• Removed secondary runtime caps that could override the Sampler configuration after the initial policy was applied.

Native H3 compute passthrough

Added a true native compute path for high-VRAM configurations.

With:

• memory_mode = normal
• attention_mode = existing
• MLP chunking disabled
• LongMedia VRAM guards disabled

LongMedia can leave the H3 transformer block execution on the stock ComfyUI path instead of routing it through the LongMedia low-VRAM block wrapper.

This keeps LongMedia planning and runtime orchestration while avoiding unnecessary chunking overhead on GPUs with enough VRAM.

Sampler controls now affect the real runtime

Performance-related Sampler parameters now preserve the distinction between:

• the value requested in the UI
• the effective memory policy
• the actual runtime configuration

This makes manual performance tuning meaningful and allows proper benchmarking between chunk sizes and the native path.

Reporting improvements

Fixed misleading runtime reporting around attention and memory execution.

The report now more accurately reflects:

• requested vs effective MLP chunk size
• whether MLP chunking is actually enabled
• native block passthrough state
• actual attention mode
• Sol execution state
• active VRAM guard configuration

In particular, an inactive Sol configuration is no longer reported as if embedded Sol-Attn were the active execution path.

Suggested high-VRAM benchmark

For the same model, seed, resolution, duration, and step count, compare:

8192 → 16384 → 32768 → 65536 → 131072 → 0

where 0 represents the unchunked/native MLP path.

v0.6.41 — Refiner AV Device Handoff Fix

Fixes a device mismatch in monolithic Stage 2 refine that could cause pack_latents to fail when refined video remained on CUDA while Stage 1 audio was offloaded to CPU.

The refined video is now restored to the Stage 1 storage device before AV repacking, while Stage 1 audio remains an exact passthrough.

Also adds AV device-contract validation to catch invalid mixed-device latent states earlier.

No workflow, UI, sampling, or audio behavior changes.

v0.6.40 — Director, Refiner, TAKEs & VRAM Stability

ComfyUI-MiniMax-H3-LongMedia v0.6.40

Major update since v0.5.40 focused on production usability, long-form generation, refinement quality, and runtime stability.

Highlights

• New LongMedia Director with timeline editing, MAIN/CAMERA/AUDIO tracks, per-shot prompts, FIRST/LAST anchors, Program Monitor, Ripple editing and resolution control.
• Native H3 routing for T2V, FL2VA, REF2VA, hybrid references and video edit.
• Improved segmented and multiclip generation with safer latent handoff and fewer repeated-motion artifacts.
• Reworked Latent Hi-Res + Refiner pipeline with internal Stage 2 sampling.
• New Windowed Refine switch:
  • ON — lower VRAM usage for long/high-res generations.
  • OFF — full-latent refine to avoid window transition artifacts.
• Refiner now works as a video refiner only — Stage 1 audio is preserved exactly.
• New TAKE system with full Director-state snapshots, restore, duplicate, rename and project folders.
• Major fixes for camera timing, motion context, FIRST/LAST geometry and H3 conditioning.
• Faster H3 VAE decoding with batched spatial tiles.
• Safer repeat-Queue behavior with improved VRAM/cache cleanup and OOM handling.
• Persistent Director UI layout and panel sizes.
• Full English and Russian documentation.
• Release package cleaned to the three maintained workflows:
  • Director
  • Latent Upscale / Detailer
  • SAFE 1080p 15s

v0.6.40 is the new recommended baseline for LongMedia workflows.

v0.5.40 — Semantic Modes, Native AV Editing & Memory-Safe Two-Pass H3

LongMedia 0.5.40 is the release consolidation of the development line since v0.4.40.

The public Setup model is now split into independent semantic controls for H3 conditioning, timeline construction, duration ownership, and audio behavior. This makes video_ref_edit, redubbing, MultiClip, segmentation, reference conditioning, and duration extension composable instead of being hidden behind one legacy workflow selector.

Highlights

• Semantic Setup: control_mode, h3_mode, timeline_mode, and always-visible duration_source, with legacy workflow migration.
• Planner + Cameras: stable clip identity, drag/reorder, presets, per-clip cinematography, and continuity-aware transitions.
• Native video_ref_edit: paired source AV preservation, arbitrary new-dub lip-sync, independent duration ownership, multiple Audio references, and audio-reactive prompting.
• Multi-audio ownership: Audio1 owns the preserved/source soundtrack in video_ref_edit; Audio2/Audio3 remain independent H3 conditioning references.
• Two-stage H3: denoised-x0 learned latent upscale plus optional independent high-resolution same-seed H3 refinement.
• Reconstruction + Loop Closure: native Ref2VA source editing, detail recovery, and latent macro-state loop return.
• FastH3 / FastVideo VSA compatibility: strict structural validation, runtime isolation, and portable learned-VSA execution.
• 16 GB / Windows reliability: Triton TinyCC bootstrap, exact Comfy Kitchen query streaming for impossible full QKV, preview isolation, repeat-run memory boundaries, bounded pinned RAM, and guarded native INT8 prefetch.
• Documentation refresh: current modes, audio contracts, two-pass sampling, sampler/VRAM guidance, and sanitized example workflows.

Upgrade Notes

Existing workflows remain supported through legacy node class IDs and workflow_mode migration.

New workflows should use the semantic Setup controls documented in docs/MODES_GUIDE.md.

Keep ComfyUI Dynamic VRAM enabled for oversized H3 checkpoints.

Package Identity

• GitHub: vizart-vj/ComfyUI-MiniMax-H3-LongMedia
• Comfy Registry PublisherId: noise

See docs/RELEASE_NOTES_0.5.40.md for the full changes since v0.4.40.

v0.4.40 — Continuity, Stability & UI State Fixes

A stability-focused release for LongMedia with cleaner MultiClip/segmented continuation routing, persistent UI state, sane production defaults, and fixes for latent-hires behavior when disabled.

Changes:

Fixed segmented_continuation accidentally entering the native MultiClip decode path
Improved continuity handling across clip boundaries
Hi-Res OFF no longer requires an upscaler model
Setup/Sampler settings now persist correctly after reload and workflow-mode changes
New defaults: segment_duration = 5s, manual_duration = 10s
Refine defaults aligned to the production 6 + 2 execution path

v0.4.30 — MultiClip Prompting & KJ Preview Fixes

v0.4.30 — Changes since v0.4.2

MultiClip prompt workflow

Added separate Global Prompt and Multiple Clips Prompt inputs to MiniMax H3 Long Media Planner.
Added editable per-clip prompt fields directly inside the Planner.
Added Import Prompt for manually parsing a structured prompt into individual clips.
Structured prompts using both clip_1:, clip_2:… and shot_1:, shot_2:… are supported.
Import automatically creates the required number of clip entries when the incoming prompt contains more clips than currently visible.
Imported clip prompts remain fully editable after parsing.
Added independent Auto Import behavior; pressing Import Prompt no longer unintentionally enables Auto Import.
Fixed prompt import when the structured prompt is supplied through the connected Multiple Clips Prompt input.
Existing per-clip duration and seed controls remain independent, allowing each generated clip to have its own timing and seed.

KJNodes Model Preview Override compatibility

Restored compatibility with KJNodes Model Preview Override during LongMedia / MultiClip sampling.
Reworked the sampling integration to respect ComfyUI's expected CFGGuider / outer-sample wrapper contract.
Preview Override can now receive and display sampling previews instead of remaining empty.
The compatibility path was implemented without changing the custom node folder name or breaking existing workflows.

UI fixes

Fixed Long Media Setup becoming unnecessarily wide when switching to MultiClip.
The node now preserves the user-defined/original node width instead of forcing an expanded layout.
Fixed the Config UI regression introduced during development of the new MultiClip prompt controls.
MultiClip-specific prompt controls remain in MiniMax H3 Long Media Planner rather than being incorrectly added to Long Media Setup.

Release consistency

Release version standardized as 0.4.30 across package/version metadata to prevent version mismatch.
Updated release documentation to reflect the actual Planner-based MultiClip prompt workflow.