SYSONLINE
RELv3.5.0
RACK9 PIPELINES
VRAM4.0-23.8 GB
VISION STUDIO v3.5.0 · FLUX.1 DEV + SCHNELL · SD 3.5 LARGE + MEDIUM · SDXL · SD 1.5 · LTX VIDEO · ANIMATEDIFF V1-5-2 · STABLE VIDEO DIFFUSION XT · LORA + CONTROLNET + IP-ADAPTER · CUDA 13.2 · METAL · LOCAL · FIRST
UTC10:25:37
S/N · VX-00001
MOD · DOWNLOAD · 01S/N · VX-DIST-3500ARMED
PWRWINMACLINUXSRC
LOCAL · INSTALL · YOUR · GPU · YOUR · MACHINE
FREE·MIT LICENSED·LOCAL FIRST

Download Vision Studio. Free.

One open-source desktop app for local AI image and video generation: FLUX.1, SD 3.5, SDXL, LTX Video and AnimateDiff on your own GPU. No subscription, no usage caps, no account, and no cloud unless you add your own key. Grab the build for your platform and you own the whole pipeline.

FREE · $0MIT LICENSEWIN · MAC · LINUX
VX-DIST · v3.5.0
S/N · VX-DIST-3500 · BLD · 2026.11.12
MIT · © 2024–2026 · ROCKY ELSALAYMEH
v3.5.0
LATEST · RELEASE · STABLE
LICENSE
MIT
PRICE
$0
BUILDS
3 OS
ENGINELOCAL GPU
ACCOUNTNONE
CAPSUNLIMITED
TELEMETRYNONE
READ · MIT · LICENSE
MOD · RELEASE · 02v3.5.0 · 12 NOV 2026SHIPPED

What changed in 3.5.0.

The first code-signed Windows release. The canvas becomes a real editor with Clone Stamp, Heal, Shape and Pen; the engine moves to PyTorch 2.14.1 on CUDA 13.2 with RTX 50 support; the boundary between the app and its engine is locked down; and jobs survive a crash. 3.5.0 raises the GPU and macOS floors, so read the first two rows before you upgrade.

CHANGED

Windows builds are code-signed.

Installers are signed through Azure Artifact Signing with the publisher Sharif Elsalaymeh, and the updater installs only builds signed by that publisher. A found update downloads only when you choose Download update. Windows 3.4.x installs were unsigned and cannot update themselves, so install 3.5.0 from this page once; later releases install from inside the app. macOS builds are not notarized yet.

CHANGED

A newer engine, and a new GPU and macOS floor.

PyTorch 2.14.1 on CUDA 13.2 supports NVIDIA cards from the RTX 20 and GTX 16 series through the RTX 50 series, on driver 580 or newer. A GTX 10-series or older card now runs on the CPU, and the app names the card and the reason instead of failing on the first render. macOS needs version 14 or later on Apple Silicon. TensorRT now ships in the Windows and Linux builds, off until a model family passes verification.

ADDED

Clone Stamp and Heal retouch the image itself.

Alt+click a source and paint to clone it, with the source following the brush. Heal blends texture into the colour and light around each stroke so the patch has no seam, and with no source it finds a matching patch nearby on its own. Each stroke is an undo step and is part of the exported PNG.

ADDED

Shape and Pen draw vector layers.

Rectangles, ellipses, lines and arrows (Shift for squares, circles and 45-degree lines), and open or closed Pen paths, straight or smoothed. Each is its own layer with outline and fill that selects, restyles, moves, scales and rotates like text.

ADDED

Adjustments, filters, undo and export are real.

All fifteen adjustment sliders and the look filters now run on the pixels, so the canvas shows what you export. Undo and redo cover every canvas edit, up to 100 steps, and Export saves the flattened result to your library. A new Tool options panel gives the brush its size and colour.

SECURITY

The app and its engine are locked down.

The app and the local engine prove each other with an HMAC handshake at launch, and every call between them carries a session token, so another program on your machine cannot drive your GPU or read your outputs. The interface can reach files only inside your output folders and the files you pick, every channel to the app core is typed and validated, and a strict Content Security Policy blocks inline scripts.

FIXED

Jobs survive an engine crash.

Generation jobs are kept in the local database, so a restart keeps their history and results. A crashed engine is restarted for you, and closing the window while work is running asks first.

FIXED

FLUX, FLUX Fill and SD 3.5 render correctly on your GPU.

A scheduler swap replaced their flow-matching scheduler with an Euler one, which broke every local run of those models. Swaps now start from the scheduler the model loaded with. A fresh install also targets an installed model on its first Generate.

FIXED

Hosted routes never bill a job twice.

A timed-out OpenRouter or Hugging Face generation is no longer re-sent, and local job submits go out exactly once. Cancelled Hugging Face jobs stay cancelled, and API keys never appear in logs or error messages.

SECURITY

Every high-severity advisory is cleared.

Electron 45.3.1, axios, zod, fast-uri and follow-redirects move past their advisories, and npm audit reports zero with and without dev dependencies. The engine installs only from hash-locked requirements, and packaged builds no longer answer the reload and DevTools shortcuts.

PERF

A lighter start and a quieter renderer.

The startup bundle drops from 782 kB to 442 kB because every other view loads on demand, the saved state is written after a short pause instead of on every change, and only one model is held in VRAM at a time.

Previously in 3.4.1
SECURITY

The js-yaml override now sits on the line the patch actually shipped on.

3.4.0 cleared the advisory by substituting a different major of js-yaml into electron-updater. The fix had shipped on the line electron-updater already declares, as 4.3.2, so the override is now a floor on that line rather than a major swap sitting under the code that parses the update feed. The tree resolves one js-yaml, 4.3.2, and npm audit reports zero both with and without dev dependencies.

FULL · CHANGELOG ↗Released · every entry is code in the public repository
MOD · PLATFORMS · 033 BUILD TARGETS · x64 · ARM64READY

Pick your installer.

One codebase, three native builds (Windows x64, macOS on Apple Silicon, Linux x64), each delivered straight from the Vision Studio update host with the whole engine inside. Generation runs offline once your models are downloaded. Windows builds are code-signed; macOS builds are not notarized yet, so that card notes the one-time prompt macOS shows.

PKG · WIN-X64WIN 10 / 11 · x64
WINDOWS
WEB INSTALLER · PORTABLE ZIP
.EXEWEB INSTALLER
.ZIPPORTABLE · NO INSTALL
SIGNEDPUBLISHER · SHARIF ELSALAYMEH
0.93 MBWEB INSTALLER · PULLS 3.79 GB
PKG · MAC-ARM64MACOS 14+ · ARM64
MACOS
APPLE SILICON · M-SERIES
.DMG · arm64APPLE SILICON
FIRST OPENNOT NOTARIZED · OPEN ANYWAY
586 MBFULL DISK IMAGE
PKG · LINUX-X64UBUNTU 22.04+ · x64
LINUX
APPIMAGE · x64
APPIMAGEPORTABLE
FIRST RUNCHMOD +X
6.75 GBSELF-CONTAINED · ENGINE INSIDE
MOD · REQUIREMENTS · 04SPEC SHEET · DOCUMENTEDPUBLISHED

What it takes to run.

Minimum gets the app running, on the CPU if it has to. Recommended is what the larger image and video models need. The engine uses one GPU: the first NVIDIA card (RTX 20 / GTX 16 series or newer, driver 580+) through CUDA 13.2, or Apple Silicon through Metal. Older NVIDIA cards run on the CPU.

REQUIREMENT
MINIMUM
RECOMMENDED
OS
WIN 10 · MACOS 14 · UBUNTU 22.04
WIN 11 · MACOS 14 · UBUNTU 24.04
GPU
NONE · CPU (SLOW)
NVIDIA RTX 20+ 8 GB · DRIVER 580+ · APPLE M-SERIES
RAM
8 GB
16 GB
DISK
15 GB FREE FOR THE APP
260 GB FOR THE FULL STARTER SET
MODELS
~2-24 GB EACH
STARTER SET 34 MODELS ~247 GB
INTERNET
INSTALL · MODEL DOWNLOADS
INSTALL · MODEL DOWNLOADS
GPU · VRAM · LADDERVRAM · SCALE · 24 GB
RTX 409024 GB
BEST
A+
GRADE
RTX 408016 GB
GREAT
A
GRADE
RTX 407012 GB
GOOD
B
GRADE
RTX 306012 GB
GOOD
B
GRADE
RTX 20606 GB
FAIR
C
GRADE
CPU ONLYSYS RAM
SLOW
D
GRADE
RATINGS · FROM · THE · PROJECT · NOT · A · BENCHMARK · THE · FOUNDRY · CHECKS · YOUR · CARD
MOD · FIRST RUN · 05ONE-TIME SETUPLOCAL READY

Online to set up. Local to create.

The installer already carries PyTorch and the rest of the engine. What you download afterwards is model weights, and only the ones you choose: a one-click starter set of 34 models (about 247 GB) or one model at a time from the Model Foundry. Once they are on disk, generation needs no connection, no account and no cloud.

01
INSTALL

Run the build for your OS; the engine is inside. No command line. Windows builds are signed; on macOS, allow the app once when it asks.

02
FIRST LAUNCH · MODELS

The engine starts with the app, which can take a few minutes the first time. Then install the starter set or pick models in the Foundry.

03
GENERATE OFFLINE

Render on your GPU with no network. Add models any time. On Local routes nothing leaves your machine.

ALT · BUILD · FROM · SOURCE/MIT · CLONE · & · RUN
$git clone https://github.com/Git-Rocky-Stack/vision-studio
$cd vision-studio && npm install
$node scripts/setup-dev-backend.cjs
$npm run dev

Node.js 22.22.2 or later (24.15+ and 26+ also work) and Python 3.12, the version the hash-locked requirements are built for. The setup script creates backend/venv with the release build's own install steps: the pinned PyTorch 2.14.1 (CUDA 13.2 with an NVIDIA GPU, the CPU build without one, Metal on Apple Silicon), the hash-locked runtime, and TensorRT on Windows and Linux, then runs the release self-check. On Windows, setup-python.bat runs the same script; CONTRIBUTING.md has the manual steps. npm run dev starts the Electron app and the engine, using that venv.

OPEN · REPOSITORY▼ DOWNLOAD · INSTALLERSTACK · ELECTRON 45 · REACT 19 · PYTORCH 2.14.1 · CUDA 13.2 / MPS
COMFYUI · OPTIONAL/The bundled FastAPI + Diffusers engine runs on 127.0.0.1:8000. A ComfyUI server found at 127.0.0.1:8188 at launch, or at COMFYUI_URL, takes plain image and video jobs instead.
VIEW · ON · GITHUB
VX-DIST · v3.5.0
S/N · VX-DIST-3500
MIT · OPEN SOURCE · LOCAL ALWAYS