MiniMax-H3生成管理器

"MiniMax-H3生成管理器" (Generation Manager) is the pack's everything-node: it fuses what would otherwise be a fifteen-node subgraph - the per-shot prompt handling, the scene/video/audio reference routing, the official H3 ImageToVideo / ReferenceToVideo encoding, the sampler, and the VAE decode of video plus audio - into a single node that chews through a whole script and produces per-segment clips. It's the node you reach for when the "short drama" workflow is the whole point, not a rabbit hole you want to assemble by hand.

How it works

Under the hood it runs the same encoding pipeline as the pack's individual nodes: it reads the manager's configuration (assets, model choices, generation params, sampling/decode settings) from a modal dialog, clears and repopulates a global asset pool, splits your prompt_input on [SHOT_START]…[SHOT_END] blocks, and for each segment picks an encoding path from mode:

• 纯文本生成音视频-T2VA - text only.
• 首帧图生成音视频-I2VA / 尾帧图生成音视频-L2VA - first/last frame image.
• 首尾帧生成音视频-FL2VA - first and last frame (default).
• 音视频生成音视频-VA2VA - video-to-video with paired audio.
• 多参考生成音视频-REF2VA - the multi-reference ref2va path.

Each segment's scene/video/audio instructions are resolved against the asset pool by name (it also honors @assetname mentions in the prompt), then encoded, sampled, and decoded to image frames plus audio. Results land in an internal bus pool keyed by segment index.

The inputs that matter

• mode - the six-way generation-mode switch above. This decides which official H3 encoder path runs.
• video_count - number of segments, 1–12.
• prompt_input - the script/story, with [SHOT_START] blocks and optional @ asset references.

Optionally it accepts clip, vae, audio_vae, and model directly. If those aren't connected, it can't generate - it'll still do the segmentation and report errors per segment, which is a handy dry-run behavior. Outputs are 生成总线 (a JSON bus describing each segment: mode, prompt, has_image, has_audio, frames, errors) and 资产清单 (the asset manifest). The bus feeds the pack's "视频保存分配" (Video Save Distributor).

How to install it

Part of the whole pack - ComfyUI Manager, search "ComfyUI-JZL-MiniMax-H3", or:

cd ComfyUI/custom_nodes
git clone https://github.com/wjluoxiao/ComfyUI-JZL-MiniMax-H3

Restart. Heavy dependencies here: torch, torchaudio (audio VAE encode), and imageio-ffmpeg as a fallback for pulling frames out of reference videos when your system lacks ffmpeg. It expects a recent ComfyUI with the official H3 nodes available (comfy_api.latest), and the audio_vae is required for the VA2VA/REF2VA modes.

Common issues

Two things bite people. First, the modes have prerequisites: REF2VA/VA2VA refuse to generate without audio_vae, and if clip/vae/model aren't wired the node silently degrades to "segmentation only" - check the bus JSON for error entries when nothing comes out. Second, asset names in the pool must match the scheduling instructions exactly (the pack uses names like 图片1角色孙悟空); a typo means an empty slot and a clip that ignores its reference. And the usual H3 caveat stands: the ~42.5GB weights come from ComfyUI's official H3 loaders and the MiniMax H3 Community License excludes the US, EU, UK, and South Korea from local runs.