H3 Studio is no longer just a one-clip box. Director → Cinema Studio is a short-film desk: you import a text package, the cards fill, then Produce builds reference stills first and scene clips after. Everything runs locally — ComfyUI + MiniMax H3, NVIDIA GPU.
Pinokio: Update → Stop → Start, then Ctrl+F5 in the browser.
Two paths
Seamless (~1 min)
Feels like one take: last-frame / Multishot chain. If character stills are on, Studio turns seamless off. Face lock and a single unbroken take do not run together.
Character / location
Cards + a shot list. When a name appears in a shot, that card’s images become H3 refs. This is the real film path.
Text package (JSON)
Schema: h3-cinema/v1. Another AI can write the story; you Import. No images, no URLs — text only.
All visual fields must be English (H3 requirement): looks, action, camera, place notes.
Dialogue may stay in the story language (dialogue_lang: tr / en / …).
Import does not start the queue. Cards sit ready. Work starts when you hit Produce.
Draft = 480p / 10 steps. Final = 720p / 20 steps.
Template: studio/schemas/h3-cinema-v1.example.json
Cards
Character — name, English appearance notes, optional voice / LoRA.
Creature — a full non-human creature (not a person, mask, or cosplay).
Location — one establishing frame.
You can drop in your own photos. If a card has none, Generate visual or Produce builds the stills.
Reference stills (important)
H3 face lock does not want a collage. The old “one PNG, three panels” is no longer sent as a single ref.
Character / creature: a short silent sheet clip is generated (one frame, same identity locked).
The video is deleted; the last frame stays.
Studio crops that frame into three PNGs: portrait · full front · back.
The card and Comfy keep three files.
If the shot names the person, H3 gets those files — portrait first, then front/back if slots remain. Max 9 refs.
Locations are not split. One cinematic still.
Older films that still had a collage (e.g. Kaldrun) were split once. New produces do steps 1–5 automatically.
Produce order
One GPU job at a time:
Character stills
Creature stills
Location stills
Scene clips (t2v or continue)
t2v = new clip / hard cut.
continue = previous last frame. If the cast does not overlap, Studio can force t2v.
Audio
Dialogue is real speech in [Turkish] … or [English] … only.
Sample tags (the word line, etc.) are not put in the prompt, so H3 does not read them aloud.
Same name, same voice. Score is mixed later; H3 does not generate BGM.
How to use it
Open Director.
Choose Character / location.
Import JSON, or write cards and shots by hand.
Pick Draft or Final.
Hit Produce. Stills finish first; scenes queue themselves.
Gallery, optional Telegram (finished clips as video).
Zip / export.
The Scene tab is still one-clip (new / continue / ref / face / v2v). Film work lives under Director.
Also in this update
Start no longer dies when git pull prints error: on a dirty tree.
Support: GitHub Sponsors · Open Collective. After approval and the first public donation, names scroll on the thank-you ticker.
Repo: https://github.com/erdinoral/minimax-h3-studio
