ComfyUI Node
H3 Story Planner
A ComfyUI node in H3 Planner with 25 inputs and 3 outputs.
H3 Story Planner
- project
- cast
- timeline
- beat_sheet
- report
◄idea►
◄total_seconds30.0►
◄min_segment_seconds5.0►
◄max_segment_seconds10.0►
◄formatcinematic ad►
◄dialogueauto►
◄audio_roleperformed on camera►
◄window4►
◄single_call_max8►
◄chain_first_framesfalse►
◄providerOllama (Local)►
◄ollama_urlhttp://127.0.0.1:11434►
◄ollama_modelqwen3-vl:8b►
◄temperature0.35►
◄reuse_existingtrue►
◄seed0►
◄style_prefix_override►
◄api_key►
◄api_model►
◄max_output_tokens8192►
◄num_ctx16384►
◄keep_alive10m►
◄request_timeout900►
CategoryH3 Planner
Inputs (25)
| Name | Type | Default | Description |
|---|---|---|---|
| project | H3_PROJECT | — | |
| idea | STRING | the whole brief: what happens, who is in it, what it is for | |
| total_seconds | FLOAT | 30.02–600 | — |
| min_segment_seconds | FLOAT | 5.00.5–15 | shortest clip the planner may ask for. Longer clips cost less per second of video, because the fixed per-run overhead is amortised. |
| max_segment_seconds | FLOAT | 10.01–15 | longest clip your card can render. A ceiling, not a target — the planner varies length to suit each beat. |
| format | COMBO | cinematic ad | 7 options: auto, short film, cinematic ad, ugc, product, explainer, +1 |
| dialogue | COMBO | auto | whether anyone speaks on camera. 'spoken lines' makes the planner write the actual script and render it as H3 <d>[English] ...</d> dialogue with speaker IDs; 'auto' decides from the brief; 'none' keeps it picture only. |
| audio_role | COMBO | performed on camera | what the connected audio IS. The first three treat it as the soundtrack, reused exactly: 'performed on camera' when a visible subject raps or sings it. 'voice sample' is the opposite and is the one for an ad: the audio is only a timbre, the dialogue is written here and generated in that voice. Ignored with no audio in the cast. |
| window | INT | 41–12 | segments written per prose call once the piece is too long for one reply. Each window sees the previous window's closing state. |
| single_call_max | INT | 81–40 | at or below this many segments the whole video is written in ONE prose call, which is what makes it consistent. Above it, windows are used. |
| chain_first_frames | BOOLEAN | false | start each segment from the previous one's last frame, within a scene. Hard visual continuity at the joins, but artifacts compound across a long chain — leave off unless you need it. |
| provider | COMBO | Ollama (Local) | 1 options: Ollama (Local) |
| ollama_url | STRING | http://127.0.0.1:11434 | — |
| ollama_model | STRING | qwen3-vl:8b | — |
| temperature | FLOAT | 0.350–1.2 | — |
| reuse_existing | BOOLEAN | true | keep segments already written from the same brief |
| seed | INT | 00–4294967295 | change to replan the whole video |
| castopt | H3_CAST | — | |
| style_prefix_overrideopt | STRING | — | |
| api_keyopt | STRING | — | |
| api_modelopt | STRING | — | |
| max_output_tokensopt | INT | 8192256–32768 | — |
| num_ctxopt | INT | 163842048–131072 | — |
| keep_aliveopt | STRING | 10m | — |
| request_timeoutopt | INT | 90030–3600 | — |
Outputs (3)
| Name | Type | Description |
|---|---|---|
| timeline | H3_TIMELINE | — |
| beat_sheet | STRING | — |
| report | STRING | — |