local control desk
Makes storyboard pictures, and optional short videos, from a script. All work runs on computers that you own.
what it conducts
mok-tua doesn't replace these — it drives them. Story in, shots out, without hand-juggling five separate apps.
Node-based diffusion engine — the actual pixels get made here.
github.com/comfyanonymous/ComfyUI → cinema languageCamera, lens, and shot-grammar layer — turns "wide, low angle" into a real prompt.
github.com/cocktailpeanut/directorsconsole.pinokio.git → video studioModel-agnostic video generation front end — WAN, LTX, and friends behind one UI.
github.com/Blizaine/Maestro → app installerOne-click install/run for the surrounding AI app zoo.
pinokio.computer → package managerShared model pool + launcher, so weights don't get duplicated per app.
github.com/LykosAI/StabilityMatrix → model routingOne API surface over local and cloud language models, gated by privacy tier.
github.com/BerriAI/litellm → voice · local · v0.7.0Lightweight, CPU-capable voice-clone TTS. Generation-verified — a real call produced a real wav.
github.com/PierrunoYT/pocket-tts-pinokio → voice · local · v0.7.0Expressive, prompt-directed voice-clone TTS. Generation-verified — but confirmed GPU-near-exclusive by a real OOM, so it never runs alongside another GPU app in the same launch chain.
github.com/PierrunoYT/DramaBox-TTS-Pinokio →proof, not promises
Two proven mechanisms. These clips are LTX-2 audio-conditioned (2026-08-15). MiniMax H3 VocalLock_V3 later passed a 5/5 verse batch on the same 16 GB GPU — native AV, different graph. Dedicated DreamTalk / LivePortrait / FaceFusion tools are still unproven.
Both shots are from the same 8-clip render — one LTX-2 job given the song's real audio track (not a separate lipsync pass), 887s active GPU time / ~2h9m job elapsed, on one 16GB GPU. No cloud render, no hand-timed edit. H3 VocalLock stamp: 2026-09-12 smoke report.
cited, not assumed
Public sources behind the 16 GB PASS — weights, turbo LoRAs, Vocal Lock, and the conductor stack. Check Hugging Face sibling files before picking a precision.
Omni-modal video + native stereo audio. Official model card.
huggingface.co/MiniMaxAI/MiniMax-H3 → weights · ComfyUIINT8 pruned, nvfp4 text encoder, video + audio VAEs — the pack that ran locally.
huggingface.co/Comfy-Org/MiniMax-H3 → turbo · 4 / 8 stepH3 T2AV / I2AV / Ref2AV configs and 768p turbo LoRAs. Do not stack two on one sampler until an A/B.
github.com/ModelTC/LightX2V → Vocal Lock · PASSVocal Lock binds a vocal stem to a named subject. This is the H3 lipsync mechanism that passed 5/5.
github.com/T8mars/comfyui-minimax-h3-audio-T8 → LTX path · RESTAudio-conditioned LTX-2 jobs via /api/v1/generate. Gradio port drifts; :7860 is DreamTalk, not Maestro.
github.com/Blizaine/Maestro → orchestration · laterStructured run/monitor/introspect against a remote ComfyUI. Adopt when the GPU is idle — not mid-render.
github.com/Comfy-Org/comfy-mcp → indexCheckpoints, turbo table, community workflows. Index only — not a second source of truth.
github.com/wildminder/awesome-minimax-H3 → this lab · 2026-09-12What passed, what is still unproven, and the full cite table.
docs/reports/SMOKE_TESTED_CAPABILITIES_2026-09-12.md →the pitch, short version
Feed it a script or sides. It comes back with shot-by-shot stills, ready for review before anything renders.
Same pipeline, several open video models underneath — swap engines without rewriting the shot.
Generate a shot more than once, mark the take that worked, assemble picks into one ordered cut.
No render farm required — one GPU box is enough; a LAN full of them scales further.
planned, not shipped
Earmarked overflow tools for when the local roster is saturated or missing a capability — registered, not wired in as a dependency. Every one of these is labeled with what was actually confirmed, not what was hoped for: 2 of the 4 checked so far were down or speaking an incompatible API when tested.
Would-be workaround for a local LuxTTS install permanently blocked by an upstream Rust/tokenizers dependency deadlock. The one public Space hosting it returned a runtime error.
Well-known open voice-clone TTS, a real gap vs. the local roster — but its official Space was returning a runtime error, plausibly because Coqui AI (the company) has since shut down.
Distinct voice-clone architecture. Space is live, but its legacy websocket-queue API isn't speakable by the current gradio_client version — needs a different client, not a simple call.
Meta's official space, distinct from the local ACE-Step. Confirmed up with a modern API, but a first test call needs a real reference clip, not an empty one, to succeed.
RVC voice conversion, Stable Audio 3, LivePortrait, SadTalker, MagicAnimate, LatentSync, and an official MiniMax H3 comparison point — registered as candidates, not yet call-verified.