JZL - 🤖 MiniMax-H3短剧导演台Mini

Want to make a short drama - six shots, an ongoing character, reference images and music clips - out of MiniMax H3? The slow way is a pile of nodes: a script LLM, a reference scheduler, a ref2va encoder, a sampler, a decoder, wired by hand. The fast way is this thing. "短剧导演台" means short-drama director's console, and the Mini flavor does exactly what the name implies: story in, already-encoded shot list out. It's the front half of a one-node pipeline that a pack called ComfyUI-JZL-MiniMax-H3 ships in both Pro and Mini trims. Mini skips sampling and decoding entirely - it hands you ready-to-sample conditioning and latents and lets you own the sampler. If you prefer your pipelines transparent, that's not a bug, it's the feature.

What it actually does

Under the hood it runs a four-step chain: it loads your reference assets (images, videos, audio) into a pool, calls a local LLM to decompose your story into N shot segments, schedules the right references to each segment via @mentions and scene/video/audio slots, then runs the official ref2va reference encoding on every segment. What comes out is a list of positive conditionings and a list of latents - one pair per shot - plus a passthrough of your model and VAEs. Downstream, you hook each pair into a sampler and decode. The LLM part is the fun bit: the pack abandoned llama-cpp-python and now shells out to a llama-server subprocess (prebuilt llama.cpp b10436), so it runs, finishes, and releases VRAM instead of squatting on your GPU. It also means there's a one-time setup step before the LLM modes work, which is the thing everyone trips over.

The inputs that matter

Run mode is where you steer the whole thing. 故事拆解模式 (story decompose) splits your story into N segments with no creative expansion. 故事扩展模式 (story expand) writes the story out further first, then splits. 穿透生成模式 (passthrough) skips the LLM entirely - feed it a ready prompt, and if it contains [SHOT_START] blocks it goes segment by segment, otherwise single segment. 仅提示词输出 (prompt-only) runs the LLM and outputs text only, no encoding - handy for iterating on a script cheaply.

Three more you'll actually touch:

• 故事名称 (story name) - mandatory. Leave it blank and the node hard-stops with "必须填写「故事名称」" and blocks downstream. It's used for save paths and logs.
• 生成视频数量 (video count) - the segment count, 1–48, and it controls how many shots the LLM decomposes into.
• megapixels / aspect_ratio / duration - resolution is computed with the official MP × 1024² formula (rounded to 32), and each segment is 4–15 seconds at 24fps, frame count aligned to H3's 17k+5.

Optional model inputs are model, clip, vae, and audio_vae - all passthrough. The encoding actually needs CLIP + VAE; if they're missing it still writes the script but refuses to encode and blocks downstream with a clear message.

The outputs and the wiring

The two list outputs are the point: 正向条件 (conditioning) and Latent, one entry per segment, matched by index. For a six-shot story you get six conditioning/latent pairs to feed six samplers (or however you want to handle lists in your ComfyUI version). The other outputs are the passthrough 主模型 / 视觉VAE / 音频VAE, a Latent放大参数 float to feed a second-pass (二采) upscale node, and 已拆解剧本 - the full post-LLM script as text, which you'll want for debugging.

Install and gotchas

ComfyUI Manager (search "ComfyUI-JZL-MiniMax-H3"), or cd ComfyUI/custom_nodes && git clone https://github.com/wjluoxiao/ComfyUI-JZL-MiniMax-H3, then restart. Dependencies are light - torch, torchaudio, plus imageio-ffmpeg from requirements.txt as a video-frame-extraction fallback - but it needs a recent ComfyUI with comfy_api.latest (V3 io.Schema style). Before first LLM use, run the runtime installer in the pack dir: python install_runtime.py (Windows: double-click install_runtime.bat). --dry-run checks without downloading.

Where people get burned: the English README only documents the sibling reference-encode node, so this console is effectively undocumented in English - expect Chinese labels and to learn by doing. The MiniMax H3 weights themselves ship under a community license that excludes the US, EU, UK, and South Korea, so check your region before investing in a workflow. And if a segment has video/audio references but you didn't connect audio_vae, that segment silently skips encoding - the most confusing failure in the box.