Cinematic shorts

A 30–60 s film from a brief — directed shot list, character references, Wan 2.2 A14B shots, an upscaled cut with ramps, grade and titles, music, foley and voice — each step an ordinary Studio run, gated frame by frame.

For everyone

A 30 to 60-second film, vertical (9:16: Reels, TikTok, Shorts) or wide (16:9), from a brief: script → character references → keyframes → shots → cut → voice-over → sound. Every step is an ordinary Studio run — metered, checked by the content policy, listed in your run history — and the estimate is shown before anything runs (ADR-080).

Make one

bash

orazaka login
orazaka cinema new my-film --brief "…" --format superhero-vfx --aspect 9:16 --duration 45 --shots 10
orazaka cinema estimate my-film          # credits, time and GPU cost, before anything runs
orazaka cinema run my-film --max-runs 1  # the script only: read production.json, adjust a shot if you want
orazaka cinema run my-film               # the rest; stop any time, the same command resumes
orazaka cinema fetch my-film             # keyframes, shots and the film, to look at them
orazaka cinema reshoot my-film 3,7       # a shot you do not like: a new seed, then the cut and the sound again

Formats: superhero-vfx (original action and visual effects), social-reel, news-explainer (an original, declared AI presenter on a general topic), trailer, custom.

The steps

Step	Studio	Model	What it makes
Script	`cinematic-script`	the catalogue's chat model (or `--script-model`)	title, visual bible, characters, music, voice-over, titles, 8–12 shots (camera, action, VFX, keyframe and motion prompts, transition, slow motion)
References	`image-generation`	FLUX.2 [klein] 4B	one still per character, before any shot is paid
Keyframes	`image-generation`	FLUX.2 [klein] 4B	the first frame of each shot: the shot, every character in it described in full, the visual bible
Shots	`video-generation`	Wan 2.2 I2V A14B, 4 distilled steps	5 s at 720p from its keyframe (or from the previous shot's last frame for a continuous shot)
Cut	`cinematic-cut`	SeedVR2 3B + ffmpeg (worker)	every shot upscaled to 1080p, trims, slow motion and speed ramps, 16→30 fps motion-compensated, push-ins, LUT grade, transitions, titles in the Krizaka type, synthesized foley (ambience, whooshes, impacts), the AI label
Voice-over	`speech-synthesis`	Kokoro-82M (Apache-2.0) on the GPU host	the narration
Sound	`video-sound`	ACE-Step 1.5 (MIT)	the music bed under the cut's foley, ducked under the voice, -14 LUFS

Directed, then gated

The script Studio carries the Krizaka cinematic kit: every shot names its lens, camera move, light, depth of field, atmosphere and pace. Read the shot list before anything is rendered (--max-runs 1), and edit production.json where it is weak — the file says what a person changed. After the shots, look at every one frame by frame (hands, faces, text, flicker, objects that should not be there): orazaka cinema reshoot <dir> 3 makes one again with a new seed, or shorten the shot in the shot list to keep only its clean seconds. Nothing mediocre ships.

Characters, costumes, names and logos are original: never a studio's characters or a brand's trademarks, never a real person's name or likeness. The script's prompt says so; the content policy checks every run.
A news format has an original presenter who is clearly an AI, on a general or clearly fictional topic.
Nothing sexually explicit.
Every film says AI-generated — drawn on the picture and written in the file's metadata.
Foley needs a commercially licensed model; none is installed (MMAudio is non-commercial, Stable Audio Open is gated behind a licence acceptance), so the cut synthesizes its foley with ffmpeg — no model, no licence question.

What it costs

Every figure — time per step, GPU time and cost per film, cost per minute of finished film — is generated from the production files of the showcase films: see Measured runs. The estimate orazaka cinema estimate shows is the same per-step medians times the steps still to run, beside each Studio's own credits.

Where it runs

The shots, keyframes, cut, music and voice run on one 80 GB GPU host with the cinema profile (scripts/gpu-dev.sh runpod start, GPU_RUNPOD_PROFILE=cinema GPU_RUNPOD_DISK_GB=250): Wan 2.2 I2V A14B with its distill LoRAs, FLUX.2 klein on sd-server, SeedVR2, ACE-Step and Kokoro side by side. The script runs on the platform's chat model.

Edit this page on GitHub