MiniMax H3 Loop Plan Generator V2 🐑

MiniMax H3 makes 4–15 second clips with native audio. Any story longer than a sneeze is therefore a chain of clips, and the hard part isn't generation - it's continuity. Clip two has to pick up where clip one ended, with the same faces, the same room, the same framing logic, and the model has no memory of any of it. The Context-Loop plugin solves that by driving a Production Plan from a JSON blob, and then you have to write the JSON.

This node writes the JSON. You give it a paragraph and a scene count; it gives you a plan_json with per-scene prompts, chain-continuation rules and a shared style prefix, ready for the Production Plan node's plan_json_input.

What it actually does

It's an LLM node, not a template filler. llm_service_connector takes one of the pack's connector nodes - Ollama locally if you have the VRAM, or an API key for SiliconFlow, ZhiPu, Gemini, OpenAI, Anthropic, MiniMax itself, whatever you've configured. The default temperature is 0.4 and timeout is 300 seconds, which tells you roughly what the author expected: a lot of tokens and a slow call.

user_input is free-form: a concept paragraph, or speaker：line dialogue, or a mix. Dialogue gets extracted and packed deterministically; narration goes through the storyboard LLM. Leave it blank and you get a default concept, which is a handy way to see the output shape before committing your own idea.

scene_count at 0 means auto - derived from dialogue line count and pacing, holding at most two spoken lines per scene inside the 14-second H3 window. total_duration_seconds at 0 works the same way. The author is explicit that pacing is speech tempo, not cut density: it changes how fast people talk, not how many times you cut. People who expect it to alter shot count will be confused.

The inputs that change your footage

camera_discipline is the one I'd touch first. auto leaves the storyboard LLM free to alternate shot sizes, and it has a hard rule to cut to close-ups, so a concept line saying "fixed camera" is only a soft hint. locked is enforced in code after normalisation: one composition for the whole board, camera_movement always static_lockoff, no close-up cuts. no_closeups keeps medium and wider with free movement. If you've ever watched an LLM ignore your "keep it wide" instruction, this is the fix - it's not asking nicely.

generation_mode trades money for continuity: per_shot makes one LLM call per scene (recommended) and single_call does the whole board in one cheaper call.

reference_mode decides what the pictures are for - t2va ignores them, i2va anchors scene 1 to <Picture 1>, fl2va alternates end targets, ref2va keeps N pictures active for every scene. Wire images into images (one IMAGE batch): batch[0] is the opening frame for the First-Scene Image Gate, batch[1..] the per-scene end targets.

Then there are the sockets: character_1..9 and scene_1..9, all MIE_ASSET, auto-growing as you wire the tail. Wire a Load Asset's asset output in and the roster is built straight from the bundle - the authored description, voice line and picture, no vision caption call at all. Scene bundles are marked as environment. Slot order (characters, then scenes) is <Picture N>, and wiring two copies of the same asset raises rather than quietly duplicating.

seed_mode defaults to per_scene_increment, which derives each scene's seed from one base the same way the upstream Context-Loop plugin does - seed = 0 means clock-derived per run. Back to seed/determinism if you're comparing prompt variants.

Outputs

plan_json is the headline - string in, Production Plan's plan_json_input, and the schema is byte-compatible with the V1 generator's, so existing wiring doesn't care which one you use. summary is a human-readable digest for a Show Anything node. board_kind tells you whether the board came out as dialogue or narration. images is the assembled picture batch for the video chain. assets is the roster, which goes into Asset → Slots 🐑 so the reference images and audios are literally the ones the prompts were built from - alignment by construction rather than by careful wiring.

Install

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

Or ComfyUI Manager, search ComfyUI_MieNodes. You also need the consuming end - ComfyUI-MiniMaxH3-Context-Loop - and an LLM: either a connector node plus a key in mie_llm_keys.json, or SetOllamaLLMServiceConnector with a local model and no key at all. No model weights ship with this node itself.

Things that go wrong

character_2 is not a character asset bundle means a scene bundle landed in a character socket. The error names the socket, so believe it and move the wire.

More than nine wired bundles raises outright - the H3 manifest caps at nine pictures. And long boards mean long LLM calls: if you're seeing timeouts, raise timeout rather than re-queueing, because a re-queue pays for the whole call again.

The failure that doesn't error is the interesting one. If enhance_user_input (the enhancer) rewrites a dialogue board, it's designed to pin a paraphrase back to your exact line on the first try rather than retry - so if a spoken line comes back altered, look at your source text before blaming the model.