MiniMaxDirector 0.16.0 — beta

Beta. The 0.16 line's first public build. It is usable and the test suite is green,
but this is a lot of change at once — expect bugs, and please open an issue when you hit
one. There is nothing useful to fall back to: 0.1.0 is a different program.

0.1.0 was the only release before this one. 264 commits landed since, so these notes
cover the whole arc rather than the last week.

---

The compiler

• The full six-section format for a clip that carries references — subject_definitions,
  summary, retention_analysis, detailed_description, overall_soundscape,
  non_diegetic_music — and the plain integrated form for a clip that carries none. One
  prompt shape per case, chosen from what is on the timeline.
• Task types are worked out and combined without repeating themselves: reference
  generation, keyframe completion, video continuation, video editing, audio reuse.
• Dialogue in the exact shape the guides ask for — <d>[Language] the words</d>, with
  the speaker, their (Sx) and their delivery outside the tag, verbatim words inside it.
  A shot can hold a conversation rather than a chorus; a line can cross a cut or be cut off
  by the end of the clip; a voiceover is marked as one.
• <Subject n> separates a person from the photograph they came from, and a shot's
  tokens start a sentence of their own.
• Frame anchors define themselves by the frame they are, not by their filename, and
  take a shows box saying what is in that frame.
• The rest of the published vocabulary: the camera as motion type × amplitude × speed,
  a move that can simply be written in words, on-screen text quoted exactly, storyboard
  references, <scenetrans>, and the separate marker vocabulary audio uses.
• A file used to define something is cited inside that thing's definition — a carried
  face, a motion source, a timbre reference — instead of being given an entry of its own.

WHO & WHAT

• One card per thing the prompt has to name: people, props, costumes, places, a look. Not
  a cast list — a card is the only place a file is described.
• A card can take its face from a picture, its motion from a video and its voice
  timbre from a recording, and can be carried onto somebody else for a face swap,
  which folds it into that person rather than adding a second one.
• The voice box is the speech switch: a card speaks because it says how it sounds. A
  card can also be muted on its own.
• On the dialogue row the cast is a deck of cards that fans out when you hover it; a
  click gives the line to one face, Cmd/Ctrl adds the second.
• A card with no file wears its initial everywhere, and <Subject n> chips write the
  number into the shot's text at the caret, so it is never typed and never wrong.

Files

• One list of everything the clip carries, placed or not — drag a file onto a track to
  place it, or leave it clip-wide for a look that has to hold throughout.
• Every file carries its own resize, and the clip can take a picture's shape on
  request rather than as a side effect.
• A file that has gone missing says so everywhere, can be put back from the block where
  its picture would be, and the run is refused while the clip names a file that is not there.

The timeline

• A playhead you can pick up, blocks you can drag, snapping to every edge, multi-select
  with a panel of its own, Cmd/Ctrl+A, and a selection that survives ComfyUI's undo.
• The camera move is chosen on the block; an empty camera stays empty.
• Times are typed in frames and read in seconds, and everything past the last frame is
  washed out.

The linter and the report

• A live report panel that checks the rules the guides state outright — clip length,
  reference clips outside 2–15 s, a shot with no words, a marker that cannot do what it
  is being asked to do — and it has a node of its own beside the prompt.

The node itself

• Four tabs across the whole piece: TIMELINE, WHO & WHAT, GLOBAL, and
  IMPORT / EXPORT — the last hands the entire piece over as one JSON, and takes one back.
• Content-sized: as tall as the panel you have open, on every tab. The card list is the
  only part with a height of its own, set by its grip or by the node's corner, stored on
  the node and restored with the workflow.

The workflow

• Ships as a whole graph — examples/minimax-director.json: timeline, live compiled-prompt
  view, loaders, sampler, both VAE decodes and video out, with a Turbo switch and an
  optional Upscale branch that splits the audio off and puts it back. No third-party
  pack needed for the default path.

---

Breaking changes since 0.1.0

A workflow saved against 0.1.x will not load like-for-like. What went:

• The director's reference sockets and its keyframe inputs — references live in the
  Files list now, and a middle keyframe is a request in the prompt, not an input.
• The director's length output.
• The standalone Who & What node — the tab is the only one.
• Presets and the Cmd/Ctrl+D shortcut.

The pack is three nodes: the timeline and the two panels that read it.

Install

ComfyUI Manager, or clone into custom_nodes/, then open
examples/minimax-director.json. Needs ComfyUI ≥ 0.31.0. No dependencies beyond ComfyUI
itself.

MiniMaxDirector 0.1.0

First release.

MiniMaxDirector adds one node to ComfyUI. You lay out shots, camera moves and audio cues on a
timeline, and it compiles them into the single structured prompt MiniMax H3
actually reads — then hands the sampler its conditioning and a clip length it will accept.

What it does

• A timeline, not a prompt box. Three tracks: what happens on screen, how the camera moves,
  what is heard. Each segment carries its own text and its own span.
• Compiles to the documented format. H3's guide specifies three fields —
  integrated_multimodal_description with [Shot N] markers and cut timecodes,
  overall_soundscape, and non_diegetic_music. Camera work and timed sound are written
  inside the shot they belong to, which is the part of the prompt the model demonstrably follows.
• Valid clip lengths only. H3 accepts length % 17 == 5 at 24 fps; the node snaps to the
  nearest legal count so a run cannot fail on an arithmetic detail.
• References wired, not typed. Drop an image, audio or video on a segment and it becomes
  <Picture n> / <Audio n> / <Video n>, numbered from the graph — so the prose and the
  wiring cannot disagree.
• prompt and report outputs show exactly what was built and what the linter thinks,
  before you spend a GPU on it.

Install

Via the Comfy Registry:

comfy node install minimax-director

Or clone into ComfyUI/custom_nodes/. No dependencies beyond ComfyUI ≥ 0.30.0, where the
MiniMax H3 nodes landed.

Notes

An example graph is in examples/, and docs/GUIDE.md walks through building a clip.
Camera vocabulary follows MiniMax's own documented terms, so the model receives wording it
was trained on rather than enum values.

MIT licensed.