ComfyUI-MiniMaxH3-TimelineDirector README

The project's README from github.com/Songssx/ComfyUI-MiniMaxH3-TimelineDirector. As one markdown file: README.md.

ComfyUI MiniMax H3 Timeline Director

Start with the MiniMaxH3 All-in-One Full Timeline Director workflow: one lightweight graph covers text-to-video, image reference, audio-driven generation, video editing, character replacement, motion transfer, digital humans, and manually segmented long-video generation. It now also supports long-form digital humans and character singing: a reference image preserves identity while a locked long soundtrack is sample-sliced across GEN windows, injected into every segment's AV latent, and restored continuously in the final output.

简体中文 · English

Headline feature: lightweight unlimited-length video generation

The plugin splits any target duration into continuous segments and completes them in one ComfyUI execution: per-segment generation with one shared seed, direct continuation from the previous AV-latent tail, adaptive Drift-Control video masking, Soft AV audio continuity, overlap removal, and final synchronized assembly. Extend the result by increasing the segment count—without generic Loop nodes or duplicated sampler chains. Practical length is limited only by local VRAM, RAM, disk space, and ComfyUI execution limits.

Download the all-in-one workflow · Chinese guide · Chinese prompt specification

Lightweight unlimited-length workflow

Two directly generated, approximately one-minute examples

Both videos were produced in one plugin execution and are 52.625 seconds / 1263 frames / 24fps. Click a poster to play or download the original MP4. All media is hosted as GitHub Release assets, so it adds nothing to the plugin clone or installation size.

Finite direct-latent continuation References with a 48-frame overlap
Play example one Play example two

Creator WeCom contact card

Creator: Shi Xiongsong
Bilibili · YouTube

An editable reference-media timeline for ComfyUI's native MiniMax H3 Reference to Video workflow. It brings reference videos, paired soundtracks, fixed Guides, standalone images, and standalone audio into one compact editing surface.

Video-generation agents should read the Chinese segmented long-video guide.

Highlights

Main long-video workflow nodes

Node Purpose
MiniMax H3 Material Planner Edits media and outputs a compact H3 plan plus an ordered Omni media bundle.
MiniMax H3 Omni Media-Bundle Prompt Bridge Sends the bundle to an installed Prompt Rewriter Omni backend and returns only rewritten_prompt.
MiniMax H3 Finite Segment Sampling Expands an acyclic graph for direct-latent continuation, masking, sampling, deduplication, and assembly.
MiniMax H3 Timeline Director (Compatibility) Preserves the original all-in-one workflow and older saved workflows.

Long-video generation needs only Material Planner Segment Plan → Finite Segment Sampling. The Plan Encoder remains registered as a hidden internal node for expanded execution graphs and old workflow compatibility.

The split architecture avoids a ComfyUI dependency cycle:

Material Planner ──Omni bundle──> Omni Prompt Bridge ──rewritten_prompt──> Plan Encoder
       └────────────────────H3 plan─────────────────────────────────────> Plan Encoder

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/Songssx/ComfyUI-MiniMaxH3-TimelineDirector.git

Restart ComfyUI and search for MiniMax H3.

The source UI is English. Simplified Chinese is provided through ComfyUI's official localization system and follows the language selected in ComfyUI settings; restart or reload the frontend after changing the locale.

Requirements

No extra pip dependency is declared. The plugin uses PyAV, Pillow, NumPy, PyTorch, torchaudio, aiohttp, and imageio-ffmpeg normally included with a compatible ComfyUI installation.

Example workflows

Only the following two examples are shipped. The first is the normal generation entry point; the second adds Omni prompt expansion.

Download workflow

This is the all-task, lightweight, no-rewiring workflow. Configure only the Material Planner's media, time windows, and prompts to run:

Connect the Material Planner's Segment Plan directly to MiniMax H3 Finite Segment Sampling. With no uploaded media it creates an empty AV latent; with references it encodes each segment's assigned media. Drag GEN windows to define duration and seam overlap, reuse one Global Prompt, or enter complete per-segment prompts. The plugin handles legal-frame alignment, Drift-Control, Soft AV, overlap removal, tail trimming, and final assembly.

The plugin does not auto-segment by reference-media duration or infer whether character identity should continue. Segment ranges, overlaps, and assignments are explicitly controlled on the timeline. For character replacement, use Editable Reference in most cases.

2. Timeline planning with Prompt generation

Download workflow

This variant adds the MiniMax H3 Omni Media-Bundle Prompt Bridge, allowing Prompt Rewriter Omni to inspect ordered images, videos, and audio before expanding an H3 prompt. Use it when multimodal material understanding should precede H3 generation.

This workflow requires MiniMax-H3-Prompt-Rewriter-ComfyUI. Follow that project for model, quantization, and VRAM requirements.

Basic usage

  1. Set width, height, and generation_seconds.
  2. Add video, image, and audio files using the toolbar or direct file drop.
  3. Move, trim, or split video clips, then place the cyan range over the interval to generate.
  4. Select a purpose for each video:
    • Fixed Guide anchors the overlap at its generated-frame positions.
    • Editable Reference sends it as <Video N> without hard-locking the original subject.
    • Boundary Only anchors only the first and last overlap frames.
  5. Enable or disable paired video soundtracks as needed.
  6. Verify the reference labels at the bottom and run the connected encoder or prompt workflow.

Videos are numbered left-to-right by their intersections with the cyan range. Standalone images and audio follow their visible bin order; drag reordering immediately updates the underlying H3 order.

Segmented long-video generation

Generate long videos in overlapping segments. Use the previous segment's final shot as the next segment's opening Guide, and describe that overlap as Shot 1 before new content. When assembling segments, remove the repeated Guide interval from the later segment. See the Chinese agent guide for the full procedure.

Finite direct-latent continuation

Finite sampling carries the previous sampled AV latent tail directly into the next opening and avoids an RGB decode/re-encode round trip. Drift-Control is always active and has no user-facing mode selector. The requested overlap is aligned down to H3's legal temporal grid (for example, 24 becomes 22 and 48 becomes 39), and that same actual value drives latent carry, decoded trimming, and assembly. The mask adapts both to the aligned overlap's video-token count and to the connected sampler's sigma schedule, including accelerated 4-step and 8-step schedules. It dynamically re-noises only the disposable video prefix while keeping the seam-side latent clean. When audio continuation is enabled, the carried overlap stays exact until its final eight audio-latent ticks, where a half-cosine Soft AV mask releases it into newly generated sound. Assembly replaces the preceding audio tail with this incoming Soft AV overlap so the transition is retained in the final output. All segments use exactly the seed shown on the sampling node. The old generic-loop helper nodes and PR #15923 dependency have been removed.

Drift-Control AV is adapted from ComfyUI-MiniMaxH3-Contex-Loop under GPL-3.0. It remains experimental and is intended for same-shot long-chain comparisons.

Credits

License

GPL-3.0

ComfyUI-MiniMaxH3-TimelineDirector (Songssx/ComfyUI-MiniMaxH3-TimelineDirector)

lipsync
Searching Database…

657 ·GitHub
ComfyUI-MiniMaxH3-TimelineDirector
What is this?It's an editable reference-media timeline UI for ComfyUI's native MiniMax H3 Reference-to-Video workflow, letting you arrange reference videos, images, audio, and Guides on one timeline to drive text-to-video, character replacement, motion transfer, digital humans, and long-form video generation.
Is this a standalone app or a plugin I add to ComfyUI?It's a ComfyUI custom node plugin, not a standalone app. You install it into an existing ComfyUI setup and it adds timeline UI nodes alongside the native MiniMax H3 nodes.
How do I install it and get it running?Install it as a ComfyUI custom node, then restart ComfyUI and search for 'MiniMax H3' to find its nodes. From there, load the provided all-in-one workflow and configure the Material Planner's media, time windows, and prompts.
What models or nodes do I need already set up before using this?You need a recent ComfyUI build with the native MiniMax H3 nodes, the MiniMax H3 Ref2VA model plus CLIP, video VAE, and audio VAE, and Python 3.10+. MiniMaxH3AddGuide is only required if you use Fixed Guide clips.
Does it automatically figure out video segments and character continuity for me?No — the plugin never auto-segments by reference-media duration and never infers whether character identity should continue between segments. You explicitly control every segment range, overlap, and media assignment on the timeline yourself.
What's the headline feature compared to a normal ComfyUI video workflow?It can generate effectively unlimited-length video in a single ComfyUI execution by splitting the target duration into continuous segments that share one seed, continue directly from the previous segment's AV-latent tail, and get stitched together with overlap removal — no generic Loop nodes or duplicated sampler chains needed.
What is Drift-Control and why does it matter?Drift-Control is an always-on adaptive video masking system used during segment continuation; it dynamically re-noises only the disposable video prefix at a seam while keeping the seam-side latent clean, which keeps long multi-segment videos from visually drifting over time.
Can it generate lip-synced or singing digital humans across a long clip?Yes. You assign a character image per segment and mark one long soundtrack as 'Locked Original Audio'; the audio is decoded once, sliced by exact sample position, and injected into each segment's AV latent so lip-sync or singing stays continuous across the whole output.
Why did the author remove the old generic-loop helper nodes?They were replaced by the Finite Segment Sampling approach with direct-latent continuation, which avoids an RGB decode/re-encode round trip and a prior ComfyUI dependency (PR #15923), making long segmented generation lighter and more reliable.
What are similar projects?ComfyUI-MiniMax-H3-LongMedia, ComfyUI-ALLinONE-MinimaxH3, FreeVideo, ComfyUI-YCNodes-MiniMax-H3, ComfyUI-H3-FaceRefine, ComfyUI-MiniMaxH3Mod, LTX-Video, LongCat-Video
Which topics is it in?talking-head, lipsync, comfyui-nodes, minimax, ai-video, image-to-video, comfyui, video-editing, text-to-video, audio-processing, video-processing, workflow
Creator
Songssx
Size
0.9 MB
Stars
657
Updated
CommitOct 14Releasev0.6.0Sep 01
Age
1mo
Commits · 90 Days