Songssx/ComfyUI-MiniMaxH3-TimelineDirector

ComfyUI MiniMax H3 Timeline Director

What it is – A community‑authored plugin for the visual‑programming UI ComfyUI that adds a single, high‑level node set for building unlimited‑length video pipelines with the MiniMax H3 model. It lets you lay out a timeline of clips (text prompts, reference images, reference video, audio tracks, etc.), split the target duration into as‑many latent‑continuation segments as you like, and have the plugin stitch the pieces together automatically.

Why it matters – Generating long videos with diffusion models normally requires manual looping or repeated graph execution, which quickly introduces quality loss, audio glitches, and visible seams. This node bundle implements:

• Finite‑segment continuation – each segment continues the AV latent from the previous one using a shared seed, so the video‑latent path never leaves the model’s latent space.
• Drift‑Control masking – adaptive masks keep the latent “on‑track” across segment boundaries, eliminating flicker, white flashes, and blur.
• Soft‑AV audio continuity – the original audio waveform can be locked and re‑injected so lip‑sync and soundtrack stay unchanged while the model still drives the motion.
• Two‑stage SelfLift sampling – a built‑in progressive‑resolution sampler (≈75 % low‑res, 25 % high‑res) that runs for every segment, avoiding the quality degradation that occurs when you upscale after the fact.
• Full timeline UI – move/trim/split clips, snap to frames, assign per‑clip modes (Fixed Guide, Editable Reference, Boundary‑Only), and see a low‑res preview proxy (up to 480×270 @ 12 fps) while the high‑res generation runs.

Key components

Node	Role
MiniMax H3 Material Planner	UI for arranging media, setting generation windows, and emitting a compact “plan” that describes what each segment should do.
MiniMax H3 Omni Media‑Bundle Prompt Bridge	(optional) Sends the ordered media bundle to an external Prompt‑Rewriter Omni backend and receives a rewritten H3 prompt.
MiniMax H3 Finite Segment Sampling	Core engine – expands the graph into per‑segment sampling, applies drift‑control masks, performs the two‑stage SelfLift schedule, and assembles the final video/audio.
MiniMax H3 Timeline Director (Compatibility)	Thin wrapper that preserves the original all‑in‑one workflow format for older saved graphs.

Installation

# Inside your ComfyUI/custom_nodes folder
git clone https://github.com/Songssx/ComfyUI-MiniMaxH3-TimelineDirector.git

Restart ComfyUI and look for nodes whose names start with MiniMax H3.

Runtime requirements

• A recent ComfyUI build that already ships the native MiniMax H3 nodes (including MiniMaxH3AddGuide if you use guides).
• MiniMax H3 model files (Ref2VA, CLIP, video VAE, audio VAE) placed in the usual ComfyUI/models/… directories.
• A MiniMax H3 latent‑upscaler checkpoint under ComfyUI/models/latent_upscale_models/ for the high‑resolution stage.
• Python 3.10+, PyTorch, torchaudio, Pillow, NumPy, PyAV, aiohttp, imageio‑ffmpeg (all typically bundled with ComfyUI).

Typical workflow

1. Set the output resolution (width, height) and total generation length (generation_seconds).
2. Drag‑drop images, video clips, and audio files into the planner’s bins.
3. Arrange them on the timeline, trim/split as needed, and paint the cyan “generation range” over the interval you want the model to fill.
4. Choose a per‑clip mode:
   • Fixed Guide – anchors the overlap frames exactly as generated.
   • Editable Reference – treats the clip as a reference without locking the subject.
   • Boundary Only – only the first/last overlap frames are anchored.
5. (Optional) Enable the Omni Prompt Bridge if you want a multimodal prompt‑rewriter to modify the H3 prompt.
6. Turn on Two‑stage sampling, pick a latent‑upscaler model, and set the high‑resolution step count (e.g., 2 steps for an 8‑step schedule).
7. Run the graph – the plugin will generate each segment, apply drift‑control, blend audio, and output a single MP4.

Performance tip – The authors report a 1536×832, 29‑second video in ~10 min on a single GPU using a 75 % low‑res / 25 % high‑res schedule. Real speed depends on VRAM, model size, and the complexity of reference media.

Example assets – The repository ships two 60‑second demo videos (finite‑segment continuation and reference‑overlap) plus the full‑featured JSON workflow (MiniMaxH3全功能合一完全体导演台工作流.json). Download them from the Releases page to see the quality and seam‑free stitching.

License – GPL‑3.0 (see LICENSE).

---

This summary is based solely on the repository’s README; no additional features have been inferred.