沐阳 H3 · 导演台（全功能）

This is the pack's flagship, the node the whole thing is named after. 沐阳 H3 · 导演台（全功能） (H3Director) is a single node that organizes the entire job - storyboard, media, generation, seams, and optionally a second-pass upscale - into one runnable unit. The bundled quick-start workflow is only seven nodes: 沐阳 H3 加载器, 沐阳 H3 取模型, a KSampler select, this, and a video output. Everything else the README talks about as separate machinery - the Media Agent, the segment plan, the anchor long video, Turbo scheduling, second pass - is orchestrated from inside it.

Underneath it's not magic: the Director reuses the same existing Myang nodes as building blocks, so old workflows and the individual advanced nodes don't get replaced. But instead of hand-wiring a dozen nodes per segment, you fill in storyboards or hand the Director a script, and it builds the timeline.

Two ways to build a timeline

• Manual storyboard (source_mode defaults to 导演台分镜卡) - each shot card gets a title, prompt, duration, and its own media. Durations snap to H3's 17k+5 frame grid. One card accepts up to 9 images, 3 videos, and 3 audio files. There's a shared-media area for character images, locations, and music used throughout; media on a card belongs to that shot. Refer to items as @图片1, @视频1, @音频1.
• Agent / long-script split - with llm_enabled, an LLM builds the timeline from total duration, per-segment duration, the media inventory, and writing rules. script_fallback holds the script (or convert that widget to an input and connect an Agent myang_prompt). skill_preset (default auto) picks a writing skill for the segment-prompt structure, and skill_text is your own rules on top. With the LLM disabled, the input prompt takes the local splitting path and burns zero tokens.

task_mode picks the generation job: pure generation (no reference video), motion transfer (each segment follows its slice of the reference video - a motion-transfer card can pick its own action source, else the global ref_video is used), or video continuation. llm_service and vlm_service control which configured LLM/VLM backs the split and whether a VLM first describes the shared media so the LLM can decide what to reference.

The settings that matter most

• h3 / model / sampler - the loader bundle, the first-pass model (base or Turbo-joined both go here), and a sampler node.
• resolution / aspect_ratio / steps / denoise / scheduler / noise_seed - the generation contract; simple scheduler and denoise 1.0 are the Turbo-friendly defaults.
• context_length - the continuity window (22 default; same scale as everywhere else in the pack).
• 起始段 (start_segment) - resume support for motion transfer: 1 re-runs everything; on a resume you set the first un-generated segment and connect the previous segment's finished clip to 前段视频, whose tail only seeds the anchor (it never enters the ref2va reference channel). 前段音频 is its optional soundtrack for the audio seam.
• The whole 二采… family - second-pass mode, resolution, steps, denoise, upscaler, chunking, passes and seed strategy, all inline. 二采模型 takes the Ref2VA base model before the Turbo LoRA.
• save_segments / segment_prefix / save_raw_segments - per-segment output so a crash doesn't cost you everything.

Outputs: images, audio, the plan_json it executed, and fps.

LLM services

Open ComfyUI Settings → Myang_node → LLM Service Settings. It supports OpenAI-compatible APIs and Ollama, multiple URL/API-key routes per service, round-robin or primary-route-first selection, and cooldown/failover after rate limits and timeouts. Keys are never returned to the browser, and config lives in user/default/Myang_node/config/llm_services.json.

Install

ComfyUI Manager search "ComfyUI-MiniMaxH3-Myang", or git clone https://github.com/civilcoco/ComfyUI-MiniMaxH3-Myang into custom_nodes, restart, hard-refresh. No extra Python deps. Load example_workflows/Minimax_H3_Myang_Director_CN.json, pick your own models, media, prompts and seed, render two segments at a 22-frame context, inspect the seam - picture and sound - before scaling up. And the standing H3 license reminder: local weights are geofenced out of the US, EU, UK and South Korea.