T8 Storyboard Director Config（分镜强化·实验）

If you've used the T8 H3 or Seedance prompt enhancers for a while, you know the failure mode that isn't about prompt quality. The prose comes back beautiful - good light, camera moves, texture language - and then you render it and the shot misses the story. The reveal lands at frame one. The handover happens twice. A character reacts to something they couldn't have seen. Two people swap screen sides across a cut and it reads as a jump.

T8 Storyboard Director Config（分镜强化·实验） is a one-widget node aimed at exactly that. Not a model, not a planner, not a second request - a config object you wire into H3 or Seedance that adds a storyboard-method block to the request the enhancer was already sending.

What it actually is

A node in T8/Utilities with one input, mode, and one output, storyboard_director_config, of a custom type. That output plugs into the storyboard_director_config socket on the MiniMax H3 or Seedance 2.0 enhancer, and one config can feed both.

Two values: 关闭 / Off (the default) and 开启（实验） / On (experimental). That's the whole surface. Off, or nothing connected, and the enhancer behaves exactly as before. On, and the pack appends its experimental storyboard method (revision 1.0.0) to the existing first generation request.

Be clear about the limits, because the name is grander than the node: no planning pass, no scoring pass, no extra paid call, and it does not touch MOSS, Music, Storyboard Pack or the Long-form Planner. Wiring it anywhere else is a no-op at best.

What the method is opinionated about

The injected advice is about information, not vibes. Each required source event gets traced to the visible or audible carrier that proves it; an event may continue across cuts but must not complete twice. Audience knowledge is separated from character knowledge, so an objective insert doesn't teach someone a fact they haven't perceived. Screen left/right is something a camera can reverse without anybody moving, and holding an object isn't ownership. Velocity and aftermath carry through the cut instead of resetting.

Then the line that justifies a whole node: fixed shot counts win. A locked single shot gets "keep exactly 1 shot; do not add cuts". A fixed N gets "keep exactly N shots". Even AUTO is spelled out as retain explicit shot boundaries… AUTO is not permission to multiply shots. If your pet peeve is enhancers chopping a five-second clip into four shots because four sounds cinematic, this is for you.

It also stays inside each target's native grammar. H3 keeps its three/six fields, keyframes, speaker syntax and Relay envelope; Seedance keeps its task intent, 镜头N ordering and media labels, and never receives H3 field names or absolute timecodes. The method survives the language-repair path, and a restored cached result carries storyboard_strategy=v20_storyboard with its revision - so "restore last result" gives you the original, not a re-run under today's switch.

The honest caveat, in the author's own words

The tooltip says benefit unverified. Turning it on may cost tokens without producing a better result, and a passage that's already complete may come back unchanged. There's no rendered-video A/B behind it, only text-contract checks and loopback queue tests. The idea is credited to a pinned community storyboard skill, explicitly non-official. Refreshingly unmarketed - and your cue to test it yourself.

The pack ships paired examples (storyboard_director_h3_comparison.json, storyboard_director_seedance20_comparison.json): two branches sharing inputs that differ only on this switch. Heads up - queueing one runs both branches, two full enhancement requests against your key. Disconnect what you don't need first. For actual evidence, the full GitHub install includes an offline A/B harness (not in the registry package):

python tools/storyboard_ab.py plan --model "your model version" --provider "your channel" --output-dir "D:/sb/plan-001"
python tools/storyboard_ab.py blind --manifest "D:/sb/plan-001/manifest.json" --observations "D:/sb/plan-001/observations.json" --output-dir "D:/sb/blind-001"

It refuses empty experiments, unpaired slots, changed settings and mismatched output hashes - more rigour than most packs bother with.

Install

Same single clone as the H3, Seedance and Music enhancers:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-prompt-enhancer-T8.git

Restart ComfyUI and Ctrl+F5 the browser. Manager users: search MiniMax H3 / Seedance 2.0 / Music 3 Prompt Enhancer (T8) or run comfy node install minimax-h3-seedance-music3-prompt-enhancer-t8. No models and no extra Python dependencies for the cloud path.

Where people get burned

The node doesn't exist in your menu. This one is version-gated hard: the storyboard config shipped in pack 1.34.0 on 2026-10-18. If Manager installed an older Active registry build you'll get "success" and no node. git pull --ff-only, full restart, and check for a duplicate directory.

It won't connect to anything. The output is a custom typed object, not a STRING. You can't feed it from a text node or read it with a show-text node; it fits only the matching optional socket on H3 or Seedance, and a mismatched object raises an unsupported-schema error rather than echoing your input.

Expecting Off to do something. Off returns nothing, deliberately. When comparing runs, pin everything else - model version, provider, seed, language, director and performance settings - or you're measuring noise.

Cost, quietly. "No extra planning call" is not "free": you're adding input tokens to a request you were already paying for, and the existing correction and retry paths can still fire and bill.

It won't fix a thin story. It coordinates how information is presented in the output you get. It won't add conflict, characters, or a reason to cut. Given a brief with nothing to present, it mostly tells the model to leave things alone - the correct answer.