MiniMax H3 Loop Plan Generator 🐑

MiniMax H3 generates 4–15 seconds per call: one scene per generation, so anything longer is chained clips - and the "pray the motion carries across the cut" part is where projects die. MiniMaxH3LoopPromptGenerator exists for that part. It turns your concept into a plan_json you plug into the Production Plan node of the upstream ComfyUI-MiniMaxH3-Context-Loop pack - though note H3's open-weights licence excludes the US, EU, UK and South Korea; the hosted Hailuo API is a different thing.

What the node actually is

An LLM-in-the-graph node, not a sampler. Wire in any of the pack's Set*LLMServiceConnector nodes and that model writes the plan. No checkpoints, no VRAM, nothing downloaded.

The output is deliberately narrow: shots plus prompt_prefix, with each shot carrying id, prompt, length and seed. Steps, CFG, canvas size and continuation mode stay on the Production Plan's own widgets and never enter this JSON. Durations round onto H3's 17k+5 frame grid at 24fps, and every scene is capped at 14 seconds - a hair under the model's 15s ceiling, with a summary warning if anything was clamped.

The inputs that matter

• user_input - one multiline box, and the whole ballgame. A concept paragraph, speaker：line dialogue, or a mix. Blank gives you the built-in default concept: fine for a smoke test, useless otherwise.
• scene_count - 0 is auto: on a dialogue board it packs consecutive turns into the fewest scenes that fit the 14s window, and a speaker change does not force a cut. Set a number and speech spreads over exactly that many scenes, extras becoming silent reaction cuts. Note what pacing does not do: fast/normal/slow is speech tempo, not cut density.
• total_duration_seconds - 0 derives the budget from line count and pacing (TTS rate plus per-turn pause); without dialogue it falls back to 15s. Set it and shots stretch, clamped 4–14s each.
• seed_mode - leave it on per_scene_increment. One seed for every scene gives consecutive clips near-identical noise and a repeated motion rhythm: the classic "why does my video pulse?" bug.
• reference_mode + images - t2va is text-only and ignores the image socket. i2va anchors scene 1 to <Picture 1>; fl2va treats batch[1..] as per-scene end targets; ref2va keeps the batch live in every scene. Wire batch[0] into the First-Scene Image Gate and the rest into the Chain Frame Index Switch.

Outputs are plan_json (into plan_json_input), summary (LLM call count, token usage, warnings - also mirrored to the console), and board_kind, your cheapest sanity check: dialogue means the verbatim contract held, narration means no spoken lines were found or extraction fell over.

enhance_user_input: on runs the enhancer's rewrite inside this node first - handy for a rough draft, but any widget change then re-runs it.

Why dialogue boards are the interesting case

If your input looks like a script, the LLM never touches your lines. A span-anchored extraction pass reports each utterance plus its character offsets, and the code checks that the slice of your input matches what came back. Lines land in the plan verbatim; dialogue the per-scene model invents gets stripped. The promise: no dropped words, no split utterances, no translation. Narration boards get no such promise - there the LLM storyboard writes beats, with pacing biasing how aggressively it splits.

Then it synthesises a shared prompt_prefix (style, lighting, palette, cast sheet), writes each scene's prompt with a continuation directive referencing the previous scene's ending, and validates the plan before returning. Validation failure raises rather than handing a half-plan to a 20-minute render.

generation_mode is the cost dial: per_shot is one call per scene, best continuity; single_call writes the whole board at once when you're iterating.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/MieMieeeee/ComfyUI-MieNodes
cd ComfyUI-MieNodes && pip install -r requirements.txt

Or search ComfyUI_MieNodes in ComfyUI Manager (it's on the Comfy Registry under publisher mie). Requirements are ordinary Python packages (requests, opencv-python, soundfile, imagehash, deepdiff), nothing GPU-shaped. At runtime you need a wired connector plus the upstream Context-Loop pack: plan_json alone generates nothing. To avoid retyping keys, copy mie_llm_keys.json.example to mie_llm_keys.json; it's gitignored.

Where people get burned

• Reasoning models eat the token budget. A thinking model spends max_tokens on its own scratch-work: HTTP 200, empty content, node raises, you blame ComfyUI. The 16384 default is sized for exactly that.
• Timeouts on long boards. 300s per call is the default, and a 12-scene per_shot run is 12 sequential calls. Raise timeout to 600 before deciding it's hung.
• board_kind: narration when you wrote dialogue means the extraction pass didn't recognise your format. One utterance per line as Name：line, no paraphrasing, and don't bury lines inside prose.