Cine con IA · Escena por apartados H3

H3 doesn't want a wall of prose. It wants a structured prompt with six named sections - subject definitions, summary, retention analysis, detailed description, overall soundscape, and non-diegetic music - which is the format MiniMax's own material is written in and the one the model was clearly trained against. Writing that by hand, every render, is where most people's H3 experience goes wrong.

This node breaks the job into eight text fields and keeps the camera out of it. Its output isn't prompt text, it's a scene object: a dictionary carrying your fields under their H3 names, which the Director de camara H3 node then compiles into the finished six-section prompt and stamps a camera clause into.

Why splitting it up helps

There's a real mechanism here, and it isn't about ComfyUI aesthetics. On modern LLM-encoded models, what separate fields buy you is clean separation - a place to put each fact so it doesn't bleed into the next one. That's the same reason structured block prompts work for characters with several attributes each: when you write the outfit and the action and the lighting in one long paragraph, the model mixes up who's doing what, and reordering or separating the parts is what fixes it.

It also fixes the commonest H3 mistake, which is doubling up. If you describe a slow push-in inside the action field and select a movement on the camera node, you've given the model two camera instructions. The pack's own tooltip on accion says it out loud: what the characters do, no camera movements here.

The eight fields, and what each one becomes

• personajes_y_referencias → subject_definitions. Who's in the shot, and which <Picture N> each one maps to.
• resumen → summary. The H3-style summary, including [reference generation] where that applies.
• retencion → retention_analysis. What identity or elements each reference image is there to preserve.
• accion → the first half of the [Shot 1] line.
• lugar_y_contexto → the second half of that same line: place, objects, light, visible context.
• estilo_visual → a style line that sits above the shot.
• sonido → overall_soundscape. Ambience and diegetic sound.
• musica → non_diegetic_music.

That merge is worth internalising, because it's where your text ends up: accion and lugar_y_contexto are joined with the camera clause into one [Shot 1] line, while estilo_visual is emitted as its own line before it. Empty fields are dropped from the output entirely - you don't get empty headings - so you can fill in the two or three that matter and leave the rest blank.

Note that this node makes no prompt of its own - it packages a scene. The final text comes from the camera node, and a scene with no camera selected still compiles into a valid six-section prompt, just without a camera clause.

Inputs and outputs

All eight inputs are multiline STRINGs and all eight are required in the schema - required here means "present in the widget list", not "must be non-empty". Every one has a tooltip in the UI that says what it's for; they're in Spanish, and they're accurate.

Outputs are scene (CINECONIA_H3_SCENE) and info (STRING). scene has exactly two legal consumers: the camera director, or the simpler prompt node's sibling path if you'd rather skip the camera step. It is not a string, so don't try to feed it to a CLIP text encode or a plain text node - that's what the camera node's prompt output is for. info is a short human-readable line - "Escena H3: con personajes · con accion" style - that's handy on a text preview just to confirm the node actually saw your text.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/chaLords/ComfyUI-Cine-con-IA.git

Restart ComfyUI and look under Cine con IA/H3 for "Escena por apartados H3". The pack installs no Python packages of its own - genuinely nothing in requirements.txt - but it does want a ComfyUI recent enough to have native MiniMax H3 nodes, since the scene you build here is only useful if something downstream can consume H3 conditioning.

Manager story: it's registered in the Comfy Registry as cine-con-ia but the versions were still pending review when this was written, so a Manager search for "Cine con IA" may come back empty. Clone it, or install the release ZIP and make sure __init__.py ends up directly inside ComfyUI/custom_nodes/ComfyUI-Cine-con-IA/.

Things that bite

<Picture N> is positional, not decorative. referencia_1 is <Picture 1> regardless of what your text says, and the Escena node further down the pipeline drops any prompt line starting with a <Picture N> whose image isn't actually connected. So if a character vanishes from your render, check the picture numbering before you rewrite the description.

Second: this is the node where putting timing in goes wrong. H3 honours the order of events in a prompt, not timestamps or waypoints - "at 2.3 seconds the camera starts to arc" is not a control, it's noise. Write the sequence in the order you want it.

And the licence that governs the model underneath: MiniMax H3's open weights are under a community licence whose Applicable Territory excludes the US, EU, UK and South Korea. Building the prompt is free; running the local weights isn't licensed everywhere.