Cine con IA · Prompt

Ask any video model for a pan and watch it shrug. Camera movement is the single most-requested and least-cooperative thing in local video prompting, and MiniMax H3 is unusually structured about it: it wants its prompt split into six named sections, and it responds to specific phrasing. Cine con IA · Prompt (CinePrompt6) is a prompt builder for H3 plus LTX-2.5, Wan 2.2, Hunyuan 1.5, CogVideoX 1.5, Mochi 1 and a free-form mode - and it ships fourteen camera recipes that were actually measured on H3 rather than guessed at.

Two ways to use it

The fast way: pick a shot from toma_h3, press the "Armar el prompt con la receta elegida" button, and answer the window it opens - who's in the shot, where they are, what's behind them, what changes in them. It writes all six sections: the subject definitions with their identity lock, the summary with its [reference generation] prefix, the retention analysis, the style line and shot, the soundscape and the music. You supply the scene; the recipe supplies the camera, the opening framing and the subject action it was verified with.

The manual way: fill the six fields yourself. subject_definitions labels each element, summary is one paragraph that must start with one of the prefixes ([reference generation], [keyframe completion], [video editing], [video continuation], [audio reuse], [audio reference]), retention_analysis is one line per label saying what's preserved, detailed_description is the style line then [Shot 1] with no timestamp, overall_soundscape is ambience and diegetic sound, and non_diegetic_music is score - or literally N/A.

Then there's camera: free text, in English, prose. The plano, angulo, movimiento and intensidad dropdowns write into it - shot size, camera height, movement, amplitude - and intensidad at normal deliberately writes nothing, because that's what the model's guide asks for. Selecting a dropdown replaces one dimension and leaves the rest of your prose alone.

The recipes, and why the conditions are load-bearing

The fourteen named shots (whip pan, snorricam, yo-yo zoom, dolly zoom, rack focus, dutch angle, and so on) are LoopForge's published camera clauses, reproduced word for word. When you pick one, its complete route is preserved in the final prompt - and several of the tested conditions are not negotiable:

• One identity plate per character, and no background plate. Describe the location in text. A background plate roughly halves the camera move.
• [reference generation], never [keyframe completion].
• 124 frames for single moves, 192 for yo-yo zoom and split screen.
• 20 steps, res_multistep / simple, turbo LoRA off. A 4-step turbo render suppresses camera motion almost entirely - this is the one setting that silently undoes every recipe.
• No timestamps or waypoints. H3 honours the order of events, not their timing.
• A person who must stay put is written standing in place, breathing softly - never completely still.

The node also warns you when a turbo LoRA is loaded in Cargar modelo, warns when your six sections still describe the previous shot, and refuses to build a prompt while a {SLOT} in a recipe is unfilled. reglas_de_oficio adds a shots-constraints block (hands keep consistent contact, nobody enters frame who wasn't established, no touching hair or face) - it's pure craft rules with no scene description in it, so it composes with whatever you wrote.

The other tabs

modelo switches the whole panel. LTX-2.5 gets ltx_prompt (one continuous paragraph, not sections), ltx_audio and ltx_negativo. Wan, Hunyuan, CogVideoX and Mochi each get their own field set with an AI-assistant recipe: copy the instruction, let an assistant interview you, paste the labelled block back, and the node spreads it across the fields and joins them in that model's order. That last part is the honest design: these tabs build text, not conditioning. They don't replace the model's loader, encoder or sampler.

Two outputs, prompt and negative, and they're independent wires. On H3 the negative output has nowhere to go - H3 doesn't take one. On Wan and LTX it does.

Install

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

Restart, then find it under Cine con IA. No Python dependencies; all the recipe buttons and the prompt-assembly window are front-end JavaScript in web/cineconia.js, so a hard browser refresh helps if a button looks stale after an update.

Traps

Copy the final prompt into a text box and read it once before rendering - the sample rates on a 20-step render are long enough that a {SLOT} you forgot is genuinely expensive. And keep the discipline the recipes assume: if you change the shot but leave the summary describing a split screen, you get a split screen. The node warns about exactly that, and it's right.