Imagination Studio Packs Video, Voice, and Image Generation Into One Local App
An open source AI film studio runs the full pipeline locally on a Mac or DGX Spark: images, video with sound, voice, and music.
- Open source local AI film studio for Mac and NVIDIA DGX Spark, Apache 2.0 licensed on GitHub
- Runs Qwen-Image-2.1, FLUX3-K, MOSS-Image, MiniMax H3, FastH3, Fish Audio Drama 3, MiniMax Music 3
- Custom TensorFold Zig runtime with int8 kernels for Metal and CUDA GB10 tensor cores
- FastH3 renders 5s 864x480 clip in 73s on Spark, 66s on M5 Ultra Mac
- Scales across up to 8 DGX Sparks, phone control via Tailscale with QR access keys
- Measured memory guards, QA gates, multi-model picture shootout, honest hardware tier matrix
Imagination Studio packages image generation, video with synchronized sound, voice cloning, music, editing, and production planning into one local application. Keys, known as drowzeys on GitHub, built it with Sam Wasserman of Wasserman Productions for Apple Silicon Macs and NVIDIA DGX Spark systems. Core generation runs on the user’s hardware without subscriptions, cloud inference, or generation API keys.
Local creative stacks often require several interfaces, model servers, and Python pipelines. Imagination Studio connects those stages through TensorFold, a Zig runtime that executes inference through Metal or CUDA. That integration keeps source material and generated assets on local hardware while giving developers one queue, library, and production workflow to inspect and extend. Optional agent integrations may still use cloud services when configured.
One app, eight work areas
The interface contains six creative engines, an Agent room, and Studio Vaults for saved assets and LoRA adapters. Together, they cover planning, generation, analysis, and finishing.
| Area | What it handles |
|---|---|
| Story | Film planning through the Wasserman Slate workflow, including casting, locations, lookbooks, shot prompts, and production references. |
| Image | Qwen-Image-2.1 Turbo and Full, FLUX3-K Fast and Full, and OpenMOSS MOSS-Image. It can also generate five-view character turnarounds. |
| Video | Clips up to 15 seconds with synchronized sound through MiniMax H3 or its distilled FastH3 variant. |
| Previs | Reference-video analysis that extracts pose skeletons, depth, edges, and a written description of the camera move. |
| Sound | Fish Audio Drama 3 supports as many as 16 speakers per take, direction tags such as [whispering], and voice cloning from a short reference. MiniMax-Music3 generates up to five minutes of 44.1 kHz stereo music with sung vocals. |
| Edit | Video upscaling and person removal using SAM 3.1 segmentation. |
| Agent room | Tool access for Claude Code, Codex, Gemini, Grok, Hermes, local OpenAI-compatible servers, and the included NVIDIA Nemotron 3.5 Bolt model. |
| Studio Vaults | A shared library for generated assets and a hub for LoRA adapters. |
Memory sets the ceiling
The installer checks available memory and disk space before downloading model weights. Features then unlock according to the machine’s unified-memory capacity, which matters because the largest video and editing models must share memory with the operating system and application.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
