Self-hosted · Open models · One GPU
Your own AI film studio.
One GPU. One click.
Block out a scene, keep your characters consistent from shot to shot, and render it with sound, up to 4K. Open models on a GPU you rent by the hour: no subscriptions, no per-generation fees, no one else's servers touching your footage.
Shots from Outrun, made in Blockbuster Studio. Not mockups.
Runs on one RTX PRO 4500 for about $0.72/hr on Runpod — you pay Runpod directly, we take nothing.
Made in Blockbuster Studio
"Outrun", 15 seconds. The camera moves and cuts were blocked out in Blender; the studio rendered every shot with sound from one reference sheet and exported the film in 4K, on a single RTX PRO 4500.
How it works
-
1
Deploy
After clicking, on Runpod pick the RTX PRO 4500 (or another card with 32 GB or more) and click Add volume (180 GB, already sized for you) — that's where your models and films live.
-
2
Create your password
Click Connect → HTTP Service and create your password on the first screen. For safety, a new studio can only be claimed in the first 15 minutes; after that, restart the pod to reopen setup. Prefer to set it up front? Create a Runpod Secret and set
STUDIO_PASSWORDto{{ RUNPOD_SECRET_studio_password }}under Set overrides. A plain value works too, but Runpod shows environment variables in plain text to anything with access to your account. -
3
Start creating
First boot downloads ~128 GB of models in the background. The studio opens in about a minute; paste a free Hugging Face token when it asks, and everything is ready in about 10 minutes, depending on the machine's network.
New to Runpod? The step-by-step quickstart walks you through it, and What you need lists everything to have ready.
Features
Films from a previs
Block out a scene in Blender and every camera move and cut lands exactly where you put it. Or start from a script.
Consistent cast
Each scene's reference sheet keeps its characters, props and location looking the same in every shot.
Video with sound
LTX-2.5 renders engines, wind, footsteps and spoken dialogue along with the picture.
Voices and Perform
Give characters their own voice, or act a shot out on camera and turn yourself into any character.
Works with AI agents
An MCP server lets Claude Code or any MCP client cast, previs, render and export a film for you.
4K export, your data
Export the finished film in HD or 4K. Everything stays on your pod's volume; nothing is uploaded anywhere else.
Models & licenses
Most of the default stack is Apache-2.0 — permissive enough to use, modify and sell what you make. LTX-2.5, the default video engine, comes with its own community license — see below.
| Model | Role | License |
|---|---|---|
| Z-Image Turbo | Fast text-to-image | Apache-2.0 |
| Qwen-Image-Edit-2511 | Multi-image editing & consistency | Apache-2.0 |
| Multi-angle LoRA | Camera angle control | Apache-2.0 |
| LTX-2.5 | Video generation, with sound | Community license |
| Wan Animate 2 | Perform: you → any character | Apache-2.0 |
| Qwen3-TTS (1.7B) | Character voices & dialogue | Apache-2.0 |
| SeedVR2 (optional) | 4K export upscale | Apache-2.0 |
| ComfyUI | Inference engine (unmodified, headless) | GPL-3.0 |
FAQ
Which GPU should I pick?
The RTX PRO 4500 (32 GB, Secure Cloud). It's what the whole workflow is tested on, including previs films and the 4K export. If it's not available, pick another card with 32 GB or more, like an RTX 5090. A 24 GB card like the RTX 4090 runs images and video, but 4K export may not fit.
Is my data private?
Yes. Everything — models, projects, generated media — lives on your pod's own persistent volume. Nothing is sent to a third-party API unless you optionally enable an LLM key for script breakdown.
What happens if I stop or terminate the pod?
Stopping the pod keeps the volume intact (you keep paying for storage, not compute). Terminating the pod deletes the volume — and everything on it — permanently.
Can my characters talk?
Yes. Give a character a voice in Cast: describe it (“gravelly, tired man in his 60s, slow drawl”) or upload a short clip of someone talking. Every line that character speaks in your storyboard is recorded in that same voice, so they sound like one person across the whole film, and the export mixes the lines in. Voices run on your pod with Qwen3-TTS (Apache-2.0).
Why does video need a Hugging Face token?
LTX-2.5, the default video engine, and its reference-sheet add-on are gated on Hugging Face. The studio's Set up video dialog links both pages: accept the terms, create a free read token and paste it in. The download starts on its own, and the token stays on your pod instead of in Runpod's settings. Video then renders with sound and spoken dialogue, at 24 fps. Its community license is free under $10M in annual revenue, but it needs a separate license from Lightricks for products that compete with theirs, and published videos must be disclosed as AI-generated.
Can I use MiniMax H3 instead of LTX?
Yes, as an opt-in. On Runpod's deploy page, open Set overrides and set DOWNLOAD_MINIMAX_MODELS to true. Shots can then render with MiniMax H3, with sound, at 24 fps (LTX-2.5 stays the default unless a shot picks H3); Perform keeps using Wan Animate. H3's community license excludes the US, EU, UK and South Korea unless you get a license from MiniMax, so check it applies to you before turning it on.
Do I need Blender or an AI agent?
No. You can make films entirely in the browser, starting from a script. Blender is for when you want to block out the camera moves and cuts yourself, and an AI agent such as Claude Code can do all of it for you, previs included. What you need lists what each way needs.
Where can I get help?
Join the Blockbuster Studio Discord to ask questions and share what you make, or read the docs.
What's the content policy?
The studio adds no content filtering of its own. The open models are only as filtered as their publishers ship them, so you're responsible for what you make and for following each model's license.