T8 Long-form Planner（长视频分段导演）

Here's the hard truth about local and cloud video models in 2026: they make great 4-to-15 second clips, and they fall apart the moment you ask for anything longer. H3 and Seedance 2.0 are exactly the 4–15s kind - so "long-form" means planning long-form, not rendering it. T8 Long-form Planner（长视频分段导演） is the planning node: it takes your total duration, splits it into contiguous segments you can actually generate, and writes each segment as its own self-contained, model-native prompt.

One LLM call does the whole job. The node computes the segment schedule locally and deterministically - you give it total_duration_seconds and segment_duration_seconds, it chunks the timeline before the model ever sees it - then hands that authoritative schedule to the LLM and asks it to fill in each segment with a full executable event chain. The model is explicitly forbidden from chaining one model's prompt into the other: every segment gets separate H3 and Seedance 2.0 variants, because they're different prompt grammars and the pack keeps them isolated on purpose.

What you set

• concept (required) - the whole-film idea or script.
• creative_brief - optional, from the Director, so identity locks and world rules survive the chunking.
• model_target - "MiniMax H3", "Seedance 2.0", or both (default).
• total_duration_seconds / segment_duration_seconds - any positive integers; 60 / 15 by default. No upper cap on the total, unlike the enhancers.
• continuity_anchors - things that must survive every segment boundary.
• transition_policy - the four options are the meat: 状态连续（推荐）/ 匹配剪辑 / 声音桥接 / 允许章节跳转 (state continuity, match cut, sound bridge, or chapter jump allowed).
• rewrite_mode, output_language, seed.

What comes out

• global_continuity_brief - the master brief for the whole piece, for keeping segments honest.
• h3_segments_json and seedance20_segments_json - per-segment native prompts for each target model, each with start/end state, continuity anchors, media bindings, and a concise prompt (the contract caps it at 360 Chinese characters / 220 English words per segment so you're not handing a video model a novel).
• handoff_table_json - the segment-to-segment handoff table, so shot N's ending state actually matches shot N+1's opening state.

Each segment prompt is short enough to feed directly into an H3 or Seedance enhancer - or into generation, if you're already working at the raw-prompt level.

Installing

Part of the comfyui-minimax-h3-prompt-enhancer-T8 pack. ComfyUI Manager: search "MiniMax H3 / Seedance 2.0 / Music 3 Prompt Enhancer (T8)". Or:

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

Restart ComfyUI, Ctrl+F5 the browser. Cloud mode needs no new Python packages. The one LLM request rides the pack's shared provider config - by default the Seedance NZ gateway; wire T8 LLM Provider Config into provider_config for the AI Workshop, an OpenAI-compatible endpoint, or local Qwen GGUF.

Honest expectations

The planner doesn't render, stitch, or edit anything - it returns text, and you do the assembling (or you hand segments to a loop/render workflow). Don't expect frame-accurate guarantees; these are prompts for generation, not shot lists for a VFX pipeline. And a lighter warning: with a big total duration and many segments, one request carries a lot of responsibility, so check the handoff table before you start generating - the plan is only as continuous as the anchor you gave it.