Questions, answered.
Vision Studio is a free, open-source desktop app for local AI image and video generation. It runs on your own GPU, with no subscription, no usage caps, and no cloud unless you add your own key. Everything you need to know before the first install, in plain language.
The index.
Tap any channel to arm it. One row open at a time, like a real console, keyboard-driven and reduced-motion safe. Every answer is checked against the app's source code, not its brochure.
Yes. Vision Studio is free and open source under the MIT license: no paid tier, no subscription, no usage cap and no account to create. If you choose to route a job to a cloud provider with your own API key, that provider bills you directly.
PRICING →Yes. The full source is on GitHub at Git-Rocky-Stack/vision-studio under the MIT license, so you can read it, fork it and build it yourself. The third-party runtimes it bundles and the models you download keep their own licenses, listed in THIRD-PARTY-LICENSES.md.
Every build carries the complete generation engine (Python, PyTorch and Diffusers) as a hash-locked bundle that checks itself at startup, so there is nothing else to install; the Windows web installer fetches that package (about 3.8 GB) while it installs. Model weights are not included. The engine starts with the app, which can take a few minutes, longest on first launch, and on first launch a setup screen offers a one-click set of 34 models (about 247 GB) or lets you skip it and add models one at a time from the Model Foundry.
HOW IT WORKS →Generation does. Once the models you use are on disk, image and video jobs run on your own hardware with no connection. You need the internet to install, to download models and for any cloud route you switch on. On Windows and Linux the app also checks updates.vision-studio-x.com for a new release 15 seconds after launch and every four hours, and downloads nothing until you choose to; set the environment variable VISION_STUDIO_DISABLE_UPDATES=1 to turn the check off.
NVIDIA GPUs from the Turing generation on (RTX 20 and GTX 16 series) through the RTX 50 series, on driver 580 or newer, through CUDA 13.2 on Windows and Linux; and Apple Silicon through Metal (MPS) on macOS. Older NVIDIA cards (GTX 10-series and earlier) have no kernels in that build, so the engine runs on the CPU and tells you why. With no GPU the engine falls back to the CPU, which works but is slow. Vision Studio uses one GPU, the first CUDA device. VRAM decides what fits: the catalog lists about 4 GB for SD 1.5, 8 GB for SDXL, 12 GB for SD 3.5 Large and 24 GB for FLUX.1, and the Model Foundry labels each model as fitting, fitting with CPU offload, or over budget on your hardware.
SYSTEM REQUIREMENTS →Image: FLUX.1 [dev] and [schnell], Stable Diffusion 3.5 Large and Medium, Stable Diffusion XL and Stable Diffusion 1.5. Video: LTX Video and AnimateDiff from a prompt, and Stable Video Diffusion from a still image. Inpainting uses your selected checkpoint, or FLUX.1 Fill for FLUX. ControlNet works on SD 1.5, SDXL, FLUX.1 [dev] and SD 3.5 Large, IP-Adapter references on SD 1.5, SDXL and FLUX.1 [dev], and you can stack LoRAs. The Model Foundry also searches Hugging Face and CivitAI for more.
ALL FEATURES →Only for gated models. Stable Diffusion 3.5 Large and Medium and FLUX.1 [dev] ask you to accept their licenses on Hugging Face and to paste a free access token into the Model Foundry. The app holds that token in memory for the session and sends it only with download requests. FLUX.1 [dev] is licensed for non-commercial use, so it is not in the one-click set; you add it yourself.
Minimum: Windows 10 x64, macOS 14 on Apple Silicon, or Ubuntu 22.04 x64, with 8 GB of RAM and 15 GB of free disk for the app. Recommended: an NVIDIA GPU (RTX 20 / GTX 16 series or newer, driver 580+) with 8 GB or more of VRAM, or an Apple M-series chip, and 16 GB of RAM. Models add up: each is about 2 to 24 GB, and the full one-click set is about 247 GB, so plan on 260 GB free to install all of it.
DOWNLOAD →By default, yes. Generation runs on your machine, your prompts, images and library stay there, the interface can only connect to the local engine, and there is no analytics, telemetry or crash reporting in the code. The app goes online for model searches and downloads you start, for the automatic update check, and for cloud routes you set up; a saved OpenRouter key also lets Settings fetch that account's model list. A cloud route sends prompt text and settings, never your images, and cloud prompt tools can include short excerpts of your earlier prompts as context (the AI Director setting).
PRIVACY DETAILS →Optionally. In Settings you can add your own OpenRouter or Hugging Face key and send prompt tools and still images to a hosted model, or video to Hugging Face. Every route starts on Local, keys are encrypted with your operating system's secure storage, and a job too large for your GPU asks before it switches to a cloud route. The provider bills you directly.
Two ways. You can import a ComfyUI API-format graph and run it on Vision Studio's own engine. And if a ComfyUI server answers at 127.0.0.1:8188 when Vision Studio starts, or at the address in the COMFYUI_URL environment variable, plain image jobs and image-to-video jobs go to that server instead of the built-in engine; text-to-video always stays on the built-in engine. That hand-off uses standard checkpoint file names and does not pass LoRAs, and image-to-video becomes the ComfyUI SVD-XT workflow, which does not receive your model, prompt or clip length. The engine reconnects in the background when ComfyUI starts or drops, and a job ComfyUI does not accept runs on the built-in engine; to keep every job there, run ComfyUI on another address.
WORKFLOWS →Windows 10 and 11 on x64 (a web installer, or a portable ZIP that runs without installing), macOS 14 or later on Apple Silicon (DMG), and Linux on x64 (AppImage). There is no Intel Mac build, because PyTorch no longer publishes macOS x64 wheels. On Windows and Linux the engine runs on NVIDIA CUDA; on Apple Silicon it runs on Metal (MPS).
DOWNLOAD →Windows builds are code-signed from 3.5.0 on, with the publisher Sharif Elsalaymeh. While that new signature builds its reputation, SmartScreen can still show a warning naming that publisher: choose More info, then Run anyway. If it says Unknown publisher instead, the file is not the one we signed, so do not run it. macOS builds are not notarized yet: on macOS 15 or later, try to open the app once, then open System Settings, go to Privacy & Security and click Open Anyway; on macOS 14, right-click the app and choose Open. On Linux, make the AppImage executable (chmod +x) before running it.
In the app's user data folder, separate from the program files: renders go to an outputs folder (Settings lets you choose a different one), downloaded models to a models folder, and the library database to data/vision_studio.db. Images are saved as PNG and video as MP4, and exporting copies the file to any folder you pick.
On Windows and Linux the app checks updates.vision-studio-x.com for a newer release 15 seconds after launch and every four hours, and Settings > About has a Check for updates button. Nothing downloads until you choose Download update, and on Windows the updater installs only builds signed by the expected publisher. Updating from 3.4.x to 3.5.0 on Windows takes one manual download, because 3.4.x builds were unsigned; later releases install from inside the app. On macOS, download new releases from this site. Every release is published on the download page with its changes listed, and your models, renders and library live in the user data folder, not in the program files.
LATEST RELEASE →Checked in September 2026: entry plans run about $10 to $15 a month (Midjourney Basic $10, Pika Starter $10, Runway Standard $15) and top individual plans $95 to $120 a month (Runway Max and Pika Fancy from $95, Midjourney Mega $120), before annual discounts. Vision Studio is $0 on hardware you already own.
DO THE MATH →