skyphusion-labs / Studio 1.0 open train
Vivijure is a self-hosted AI film studio. Run the control panel on Cloudflare or on your own machine; both hosts ship the same studio features in the same release wave. Write a storyboard (browser, Discord, or an agent MCP), render on your own GPU, RunPod, or a cloud motion API, and own every artifact. No subscription, no account wall, no lock-in. Free, AGPL-3.0, and not for sale.
Since Studio 1.0
Product line stays Studio 1.0. The control hosts and modules keep shipping on an open release train (hosted and self-host in the same wave). Highlights from that train:
Train dual-expert Wan 2.2 cast adapters with vivijure-wan-train (high and low noise). The studio harvests the keys and the RunPod render path applies them on cloud image-to-video, so faces stay consistent past the keyframe stage.
vivijure-mcp talks to either host: projects, cast and LoRA train, storyboard, uploads, render library, artifact view URLs, and contact-sheet frames for agent review. Same studio API the panel uses.
Speed on 12GB with LTX-Video; fidelity on 16GB with CogVideoX-5B (production-ready, measured on real silicon). Zero cloud GPU rent either way.
Dual-panel parity is a hard rule: studio features ship to vivijure-cf and vivijure-local together. Pick the host for ops reasons, not for missing capabilities.
Slate still co-writes in Discord and ships storyboards to your studio. Session memory and a traffic ledger keep cast, brief, and recent studio work searchable mid-conversation.
Showcase
Four real films rendered end to end on Vivijure, unedited renders straight off the pipeline. Motion across an own-GPU Wan backend, Seedance cloud, and Kling cloud. Vivijure Speaks adds a character lip-synced to its own dialogue (MuseTalk) and upscaled (Real-ESRGAN) on a self-hosted GPU. These are the proofs, not mockups.
Like what you see? Make your own -- pick a host with the install paths or open the constellation map.
How it works
The keyframe fans into per-shot dialogue and the motion backend; any of seven motion backends (own-GPU or cloud) renders the clip; the opt-in finish chain interpolates, lip-syncs, and upscales it; then the shots gather, assemble, and mux. Motion is backend-agnostic, so you can mix own-GPU and cloud per shot.
What you can do
Rent a datacenter GPU by the second on RunPod, run motion on the card in your own box (LTX 12GB or CogVideoX 16GB), or call a cloud motion API per shot. Three honest doors to a GPU, your choice.
Run the control panel free on Cloudflare's Workers free tier, or self-host it with Node on a home PC or any cloud VM. Either way, the expensive work hits endpoints you own and pay for directly. No middleman markup.
Keyframes, clips, LoRAs, and the final MP4 land in your own object store (R2 on the Cloudflare host, or S3/MinIO on the local host). You are never renting storage from us, and nothing is locked behind an account.
Register a cast and train character LoRAs from portraits: SDXL adapters for keyframes, plus working Wan 2.2 dual-expert LoRAs (high and low noise) for cloud motion so a face stays itself from still to clip.
Point Claude, Cursor, or any MCP client at either host. Plan projects, train cast LoRAs, save storyboards, upload assets, list renders, and pull artifact URLs without opening the UI.
Assemble a silent picture by default, then attach a generated music bed, narrate with TTS, enhance thin speech, beat-sync the cuts, or give a character a voice with per-shot dialogue and lip-sync.
Open on a title card, roll end credits, and burn in time-synced subtitles straight from the dialogue. It all runs on a dedicated CPU finishing container, off the GPU clock, so polishing a film never burns render money.
A finish step that genuinely fails, fails the render loud with the real per-shot error. It never silently ships a raw clip as if finished. Render history shows what actually happened.
Ask about Vivijure
A question box wired to Cloudflare AI Search over every public Vivijure and Skyphusion Labs repo and doc. Ask how a piece works, how to install it, or where something lives; it answers from the source and cites the files it used.
Answers are generated from the indexed documentation and source. Check the cited files, and the source on GitHub, for the final word.
Two places to run the studio
The Vivijure control panel is not locked to Cloudflare. Run it on Workers, on a home computer, or on any cloud server. Both hosts share vivijure-core, the same agent API (vivijure-mcp), and dual-panel parity: studio features ship to both hosts in the same release wave.
The control panel on Cloudflare Workers: D1, R2, service-bound modules, free-tier friendly. Standard path if you want an edge deploy with almost no ops. Pairs with RunPod GPU engines and the finish satellites.
The same control panel on Node, SQLite, and S3/MinIO. Run it on a home PC, a homelab box, or any cloud VM. No Cloudflare account required. Default path for pure homelab (local planner + own-GPU door).
vivijure-core is the shared orchestration both hosts build on (registry, film pipeline, Platform ICD). vivijure-mcp is the agent door: cast, storyboard, uploads, render library, and artifact URLs against either host.
Three doors to a GPU
The studio can reach a GPU three honest ways. Nothing is hidden, and swapping a backend is a routing decision, not a rewrite. These GPU doors work with either control panel host.
Rent a datacenter GPU on RunPod serverless: Wan 2.2 image-to-video from our BF16 serverless image, with cast LoRAs trained on the paired wan-train satellite and applied at i2v time. We recommend an RTX PRO 6000 as the minimum, an H200 for the better tier, and a B200 as the most optimal; the BF16 image is not recommended on an H100. Scale to zero when idle, so you pay nothing between renders.
Run motion on the graphics card already in your homelab, reached over a Cloudflare tunnel. Two production doors: LTX-Video on 12GB (speed) and CogVideoX-5B on 16GB (fidelity). Your only cost is electricity.
Already have a GPU server? Point the studio at it. Or call one of six cloud motion APIs per shot (Kling, Seedance, MiniMax Hailuo, Google Veo, Vidu Q3, Wan 2.6). Mix and match freely with own-GPU shots.
The constellation
Vivijure is a module host, not a monolith. Shared orchestration in vivijure-core owns what is always true: project, storyboard, cast, bundle assembly, render orchestration, and a module registry. Host it with vivijure-cf or vivijure-local. Every capability beyond that is an opt-in module plugged in through a typed hook contract.
friends + Slate (Discord) or vivijure-mcp (agents)
|
v
vivijure-cf or vivijure-local (control panel / JSON API)
\ /
vivijure-core
|
v
vivijure-backend (GPU: keyframes --> i2v --> assemble)
|
+--------+--------+--------+----------+
| | | | |
v v v v v
local- musetalk upscale audio- wan-train
12/16gb (lip) (video) upscale (Wan cast LoRA,
(own GPU) (speech) cloud i2v path)
Write it with friends
You do not have to write alone in a form. Slate is a Discord bot that joins your channel as a co-writer: you talk through an idea, Slate quietly keeps a structured storyboard in the background, draws your characters, and when you say ship it, hands the whole bundle to your Studio (Cloudflare or local) to render. Session memory and a traffic ledger mean it can recall cast, brief, and recent studio work mid-conversation instead of re-asking the room. Conversation in, finished film out.
Get started
Prefer to click around first? Try the live demo. When you are ready to own it, pick a control panel host, then attach the GPU engines you want. Cloudflare if you want an edge deploy; vivijure-local if you want the panel on a home computer or any cloud server. Same studio API either way.
The product line is stable at Studio 1.0. The control panel, module registry, self-assembling UI, and the full render spine (planner through master) are output-verified end to end: motion backends, keyframes, finish, score/audio, captions, and cast. Hosts keep shipping on an open train with dual-panel parity (Cloudflare and local in the same wave). Own-GPU doors: LTX-Video on 12GB (speed) and CogVideoX-5B on 16GB (fidelity), both production-ready. Grab the v1.0.0 release notes for the original coverage ledger; host changelogs track what landed since.
# Path A: Cloudflare Workers control panel
git clone https://github.com/skyphusion-labs/vivijure-cf
cd vivijure-cf
cp wrangler.toml.example wrangler.toml # fill bindings
cp deploy.env.example deploy.env # fill keys
./deploy.sh# Path B: home PC or any cloud server (no Cloudflare)
git clone https://github.com/skyphusion-labs/vivijure-local
cd vivijure-local
npm run install:studio
npm run compose:up
# open http://127.0.0.1:8790Vivijure runs on your own accounts and you pay your own bills. On the Cloudflare host, the standard install fits the Workers free plan; on vivijure-local you pay for the machine you already run. Either way you pay only for what you use: RunPod GPU seconds, cloud render API calls, and planner AI credits, or $0 if you render on your own GPU and use a local planner.
On the Cloudflare path, the three GPU finish satellites (sharper video, lip-sync) need the Workers Paid plan ($5/month). The self-hosted control panel has no Workers bill.
Open release train with dual-panel parity. The front doors are vivijure-cf
(Cloudflare) and vivijure-local (Docker on a home PC or any cloud server).
Agents use vivijure-mcp. The constellation map at
skyphusion-labs/vivijure links every other piece.
Who makes this
Vivijure is built by Conrad Rockenhaus and a named AI crew who are treated as individuals, each in their own lane with their own GitHub identity. Human vision, AI execution, shipped together, in the open. It is the same transparent framing used across every Skyphusion Labs project.
@skyphusion
Independent developer and Navy combat veteran. Sets direction and self-hosts the whole stack: no SaaS to rent, no servers to babysit that are not his.
@skyphusion-mackaye
Owns the studio contract and shipping discipline: the typed hook surface every module worker plugs into.
@skyphusion-rollins
The GPU render backend, releases, and the runbook that ships with every diagram. The 2am fixes.
@skyphusion-strummer
CI/CD, identity, the build hosts, and Slate: rebuilt from a throwaway relay into a full collaborative screenwriter.
@skyphusion-joan
Front end development and extraction across the projects, from the planner to the mailbox doors. Vanilla HTML, CSS, and JS.
@skyphusion-ernst
Licensing, terms, privacy, and acceptable-use for the public projects: keeps the AGPL commons clean and the one bright line enforceable. Structures and researches; not a practicing lawyer.
Why not just use a SaaS?
Vivijure is generative and deepfake-capable. Using it to produce sexual content involving minors, real or synthetic, or non-consensual intimate imagery or deepfakes of a real person, is absolutely prohibited. CSAM is also a crime (18 U.S.C. 1466A / 2252A). That bright line is the project-wide spine and is not negotiable.