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
UTC10:25:39
S/N · VX-00001
MOD · WORKFLOW · 01S/N · VX-FLOW-0001ROUTED
PWRINPUTGPUOUTPUTCLOUD
DEFAULT · ROUTE · LOCAL · YOUR · GPU · DOES · THE · WORK
SIX STEPS·NO SIGN-UP·ONE MACHINE

Install to export. All on your machine.

Vision Studio is a free, open-source desktop app for Windows, macOS and Linux. Your GPU runs the diffusion, and on the default Local routes your prompt never leaves your computer: no upload queue, no usage caps. Follow the signal chain from a fresh install to a finished file.

LOCAL FIRSTIMAGEVIDEOEDIT
VX-FLOW · SIGNAL · MK-I
S/N · VX-FLOW-0001 · v3.5.0
127.0.0.1:8000 · LOCAL · FIRST
IN
PROMPT
ENGINE
GPU
OUT
FILE
PROMPT
MODEL
PARAMS
IMAGE
VIDEO
EXPORT
LOCAL
SIGNAL · PATH · ON-DEVICE · LOCAL · ROUTE
MOD · SIGNAL · 026-STAGE · CHAINARMED

The signal chain.

Six stages from download to delivery. You need the internet to install and to fetch the models you pick; after that, every stage runs on your machine and your GPU unless you route a job to a cloud provider yourself.

06
STAGES · LOCAL
INSTALL
FETCH
MODEL
PARAMS
GENERATE
EDIT
EXPORT
01
SETUP
Install

Download the build for your system: Windows x64 (web installer or portable ZIP), macOS 14 or later on Apple Silicon (DMG), or Linux x64 (AppImage). The engine is inside every package, so there is nothing else to install. Windows builds are code-signed (publisher Sharif Elsalaymeh); if SmartScreen still warns while the signature builds reputation, choose More info, then Run anyway. macOS builds are not notarized yet: on macOS 15 or later allow the app once under System Settings > Privacy & Security.

WIN
.EXE/.ZIP
MAC
.DMG
LINUX
APPIMAGE
STEP 1/6
PIPELINE · POSITION
02
ONE-TIME
First launch

The engine starts with the app, which can take a few minutes the first time. A setup screen then offers a one-click set of 34 models (about 247 GB), or you can skip it and pick models one at a time in the Model Foundry. Gated models such as Stable Diffusion 3.5 need a free Hugging Face token pasted into the Foundry.

STARTER
34 MODELS
SIZE
~247 GB
OR
FOUNDRY
STEP 2/6
PIPELINE · POSITION
03
MODEL
Pick a model

Choose an image model (FLUX.1, Stable Diffusion 3.5, SDXL or SD 1.5) or a video model (LTX Video, AnimateDiff, or Stable Video Diffusion to animate a still). Stack LoRAs, place ControlNet guides or reference images on the canvas, and switch models whenever you like. The model you used last stays loaded, and switching releases it first, so only one model sits in VRAM at a time.

IMAGE
FLUX/SDXL
VIDEO
LTX/SVD
ADD
LORA
STEP 3/6
PIPELINE · POSITION
04
TUNE
Dial your parameters

Write the prompt and negative prompt, then set the frame (eight aspect ratios at 512, 768 or 1024 px, or a custom size up to 2048 px), the sampler, 1 to 50 steps, CFG and seed. Lock the seed to repeat a render, or leave it random. Every model starts from 25 steps at CFG 7.5.

STEPS
25
CFG
7.5
SEED
RANDOM
STEP 4/6
PIPELINE · POSITION
05
RENDER
Generate locally

Your GPU runs the job while a local WebSocket reports its progress, and Studio runs show the image forming step by step. Nothing is uploaded unless you have switched that job type to a cloud route. Each finished render becomes a node in the iteration tree, ready to fork or re-roll.

DEVICE
GPU
PROGRESS
LIVE WS
UPLOAD
NONE
STEP 5/6
PIPELINE · POSITION
06
DELIVER
Edit and export

Refine an image on the canvas: retouch with Clone Stamp and Heal, draw shapes and Pen paths, adjust and filter it, or run the AI tools (background removal, upscale, face restoration, fill, expand and more). Cut clips together on the timeline and export an MP4, or export any render from the library to a folder of your choice. Renders are saved as PNG and MP4 in your outputs folder.

EDIT
AI TOOLS
CUT
TIMELINE
FILES
PNG/MP4
STEP 6/6
PIPELINE · POSITION
ONLINE · TO · INSTALL · AND · FETCH · MODELS · LOCAL · GENERATION · RUNS · OFFLINE
MOD · TROUBLESHOOT · 035 KNOWN · FIXESSTANDBY

When something stalls.

The problems a first install actually runs into, and the fix for each. Every step follows what the app does in code: its out-of-memory fallback, its start-up timeout, its gated-model error and its device detection.

Out of memory, or a model marked over budget

  1. Drop the resolution tier (1024 to 768 or 512 px on the long edge) before anything else.
  2. Close other apps that are using the GPU.
  3. Choose a smaller model: SD 1.5 needs about 4 GB of VRAM where FLUX.1 needs about 24 GB. The Model Foundry shows how each one fits your hardware.
  4. The engine already retries an out-of-memory load with slicing and CPU offload before it gives up.

The engine did not start

  1. The engine can take a few minutes to start, longest on first launch. Restart the app and give it time.
  2. If the engine crashes, the app restarts it on its own; after a third crash within five minutes it stops and says so. Fix the cause shown in the backend log, then start the engine from Settings.

A model download fails with 401 or 403

  1. The model is gated. Use Accept license in the Foundry to open its Hugging Face page and accept the terms.
  2. Paste a free Hugging Face access token into the Foundry, then retry. The token is encrypted with your operating system's secure storage and kept across launches.

Generation is very slow

  1. Check that your GPU is detected in Settings, under AI & Models. With no CUDA or Metal device the engine runs on the CPU, which works but is slow.
  2. An NVIDIA GTX 10-series or older card runs on the CPU too: the bundled CUDA 13.2 build supports Turing (RTX 20 / GTX 16) and newer, and the hardware panel names the reason. A newer driver will not change that; on a supported card, use driver 580 or newer.
  3. Lower the step count or the resolution tier for drafts.

Jobs are going to ComfyUI instead of the built-in engine

  1. Vision Studio hands plain image jobs and image-to-video jobs to a ComfyUI server it can reach at 127.0.0.1:8188, or at COMFYUI_URL, and reconnects in the background when that server starts. Text-to-video always stays on the built-in engine.
  2. Quit ComfyUI or move it to another address, or unset COMFYUI_URL and restart Vision Studio.
MOD · ENGINE · 04OPEN · SOURCE · MITONLINE

Under the hood.

An Electron shell drives a local FastAPI service that runs the diffusion on PyTorch and Hugging Face Diffusers, on CUDA or on Apple Silicon's Metal. The app talks to the engine over REST on 127.0.0.1:8000, after an HMAC handshake at launch, with a session token on every call and a WebSocket pushing live render progress. All of it ships open source under the MIT license.

MIT
LICENSE · OPEN
STACK · TOPOLOGY
FRONTEND
SHELL
ELECTRON 45
UI
REACT 19
STYLE
TAILWIND v4
BACKEND
API
FASTAPI
ENGINE
PYTORCH 2.14.1
ACCEL
CUDA 13.2 · MPS
TRANSPORT
REST
127.0.0.1:8000
STREAM
WEBSOCKET
MODELS
DIFFUSERS
BACKEND · MODE

The bundled engine runs out of the box. If a ComfyUI server answers at 127.0.0.1:8188, or at the address in COMFYUI_URL, plain image jobs and image-to-video jobs go to it instead, and the engine reconnects in the background when it starts or drops. ComfyUI API graphs can also be imported and run on the built-in engine.

BUNDLED · BACKENDCOMFYUI · AUTO · RECONNECT
BUNDLED
READY
EXTERNAL
:8188
BUILT · ONBLACK FOREST LABS · FLUXSTABILITY AI · SDLIGHTRICKS · LTXCOMFYUI · NODES
GET · STARTED · FREE
VX-FLOW · MODULE
S/N · VX-04-0004 · v3.5.0
MIT · OPEN · SOURCE · LOCAL · FIRST