AI Video Studio
A local-only desktop app for AI image and video generation
powered by WanGP.
This WIP project is a fork deepbeepmeep/LTX-Desktop-WanGP.
Note: I'm more of a Solutions Architect than a coder, so I'm up-front clarifying that this is heavily coded by AI. But I have a vision of what this should be (essentially the ease-of-use of online AI Platforms, Magnific/Higgsfield etc, with the beauty of OSS and local-only AI Gen, and the power of WanGP behind it!) I can't guarantee it'll work great for everyone, so far I've only tested this on one of my systems (128gb ram - RTX 4070Ti Super) and its working great there, will try to test on other systems in due course.
- Local-Only: Powered by WanGP, no cloud/third-party/off-site generations.
- Curated models: AiVS exposes tested - fast - model profiles, not every raw WanGP model or setting.
- Simple first: GenSpace should show useful creative controls, not raw technical configuration.
AiVS is in active development. Director Mode V1 is available as a standalone workspace for planning and generating prompt-driven video timelines. Guide Audio and Control Media tracks are visible but remain locked for a later release.
- Deeper integration with WanGP, removing all cloud/third-party/non-local elements.
- Streamlined app launch & added connection indicator & refresh button for connection to WanGP backend.
- Settings:
- Added video/image output settings (Settings > Outputs)
- Added Model Manager: optional WanGP model packs, approximate sizes, status, download and cancel controls.
- Removed most other settings, but kept 'Torch-Compile' option and ensured it's hooked into WanGP.
- First run:
- Bundled Python, pip and uv automatically install the compatible WanGP GPU runtime.
- Optional model-pack selection has live transfer status and can be skipped for automatic download on first use.
- Project storage is selected at the end of setup; default:
Documents\\AiVS.
- Gallery:
- Added ability to drag and drop your own items into gallery (for easier access to regular references etc)
- Added Filtering (Type: Image/Video/Audio. Source: Generated/Uploaded)
- Added 'Bin' (folder) support so you can easily organise your assets
- Added 'List' view as an alternative to the grid views.
- Added a 'cancel' button and better indications of generation progress in asset cards.
- GenSpace:
- Prompting:
- Added seed lock button to prompt box
- added enhance prompt button to prompt box
- moved media inputs above text prompt area and made it collapsible for tidier working.
- Image Gen:
- Added Additional Image Models: Flux2 Klein 4b, Krea2 Turbo & HiDream O1, plus the existing z-image-turbo.
- Added support for input images for supporting models, including a dropdown menu to select reference type (Transfer Human Pose, Transfer Depth, Transfer Canny Edges etc):
- Z-Image-Turbo: Technically doesn't support inputs, but I've routed it so if you add one it'll use Z-Image-Turbo Fun ControlNet 6B v2.1 instead which can accept a controlnet image.
- Flux2 Klein 4b natively allows reference images, I've allows up to 5, which I think is a sane amount (not sure how many it can take).
- Krea2 Turbo currently doesn't support input images
- HiDream O1 natively allows reference images, I've allows up to 5, which I think is a sane amount (not sure how many it can take).
- Video Gen:
- Added Start/End frame support
- Added Control Video/Audio support with video trim capabilities (note when a control video is added, the 'duration' setting turns to 'auto' and is controlled by the trim length).
- Retake is visible as Retake (soon) but disabled until WanGP supports it reliably.
- Added new reframe mode with a nice easy to use framing UI that takes a control video and then uses ltx outpaint lora to expand edges.
- Prompting:
- Director Mode:
- Added a standalone workspace between GenSpace and Video Editor with the shared asset library, multiple Director timelines, contextual segment settings, preview playback and a frame-based timeline.
- Build videos from movable and resizable prompt segments, optional local/global prompts, key frames pinned to Start/Middle/End, or a frame-zero Continue Video segment. Gaps, segment reordering, ripple resizing and contextual split/delete are supported.
- Generated videos populate a Generated track automatically. Regenerations become takes, with timeline playback, scrubbing, looping, take selection, visibility and deletion controls.
- Director V1 runs at 24 fps, snaps output lengths to
8n+1, and supports timelines up to 20 seconds. Guide Audio and Control Media tracks are reserved for later releases. - See Director Mode V1 and the implementation plan for current behavior and constraints.
- Retake mode when WanGP support is ready
- LoRA UI
- Audio/TTS Gen
- Production workflow
- Windows 10 or 11, 64-bit.
- NVIDIA RTX 20, 30, 40 or 50 series GPU. GTX cards are not supported.
- NVIDIA driver 580 or newer. AiVS uses one supported runtime: Torch 2.10 with CUDA 13.0.
- VRAM: 12 GB minimum for selective/light use, 16 GB recommended for image work, 24 GB recommended for LTX video and larger model packs. This is guidance, not an enforced hardware gate.
- At least 50 GB free disk space where model packs are installed; more is needed for several packs and generated media.
The desktop installer includes Python, pip, uv, MinGit, WanGP and GPU runtime setup. It does not require Node.js, pnpm or a system Git installation.
Prerequisites:
- Windows 10/11
- NVIDIA GPU with CUDA support
- Node.js
- pnpm
- Git
- PowerShell
Recommended setup:
pnpm setup:dev:win
pnpm devsetup:dev:win prepares the backend environment, installs the WanGP GPU stack, and uses either:
- a repo-local
Wan2GP/checkout, or - an existing Wan2GP checkout pointed to by
WANGP_ROOT.
To reuse an existing Wan2GP checkout:
$env:WANGP_ROOT = "D:\Wan2GP"
pnpm setup:dev:win
pnpm devIf both .\Wan2GP and WANGP_ROOT exist, the repo-local .\Wan2GP checkout is preferred.
Linux support currently targets source/dev usage with WanGP.
Prerequisites:
- Node.js
- pnpm
- uv
- Git
- ffmpeg
- NVIDIA GPU with CUDA support
- WanGP checkout available locally or via
WANGP_ROOT
export WANGP_ROOT=/path/to/Wan2GP
pnpm setup:dev:linux
pnpm devIf WANGP_ROOT is not set, the setup script can prepare a repo-local Wan2GP/ checkout.
The backend uses Python 3.11.9, pinned by .python-version.
The Windows WanGP stack installer is:
scripts/install-wangp-stack.ps1Useful options:
scripts/install-wangp-stack.ps1 -List
scripts/install-wangp-stack.ps1 -SkipWan2gpRequirementsThe installer validates supported RTX generation and NVIDIA driver 580+, then installs pinned Torch 2.10/CUDA 13 plus curated performance wheels into backend/.venv.
Development setup and backend test commands use uv sync --inexact so normal dependency syncs update declared backend packages without pruning WanGP requirements or performance wheels installed into the same virtual environment.
Install dependencies:
pnpm install --frozen-lockfileRun the app:
pnpm devRun with debugging:
pnpm dev:debugTypecheck:
pnpm typecheckBackend tests:
pnpm backend:testFrontend build:
pnpm build:frontendFrontend validation:
pnpm validate:frontendIf pnpm tries to recreate node_modules in a non-interactive terminal, set CI mode:
$env:CI = "true"
pnpm typecheck:tsThis repo pins pnpm@10.30.3 through package.json. Use Corepack or the same pnpm version consistently to avoid node_modules reinstall prompts caused by package-manager metadata mismatches.
AiVS has three main layers:
graph TD
UI["Renderer: React + TypeScript"] -->|HTTP localhost:8000| BE["Backend: FastAPI + Python"]
UI -->|IPC via preload| EL["Electron main"]
EL --> OS["OS integration: files, dialogs, ffmpeg, process management"]
BE --> WGP["WanGP / Wan2GP bridge"]
BE --> DATA["App data, settings, outputs"]
EL --> DATA
- Path:
frontend/ - React 19, TypeScript 6, Vite 8, Tailwind CSS 4
- Main GenSpace surface:
frontend/views/GenSpace.tsx - Model profile hook:
frontend/hooks/use-image-profiles.ts - Model profile types:
frontend/types/model-profiles.ts
- Path:
electron/ - Owns app lifecycle, native dialogs, file access, export, and Python backend process supervision
- Renderer communicates through the preload bridge exposed as
window.electronAPI
- Path:
backend/ - FastAPI server on port 8000
- Thin routes call handlers; handlers call services and mutate centralized state
- WanGP bridge:
backend/services/wangp_bridge.py - Model profiles:
backend/model_profiles/profiles.py - Resolution resolver:
backend/model_profiles/resolution_resolver.py - Profile API handler:
backend/handlers/model_profiles_handler.py
| Command | Purpose |
|---|---|
pnpm dev |
Start Vite, Electron, and backend |
pnpm dev:debug |
Start with Electron inspector and Python debugpy |
pnpm typecheck |
Run TypeScript and Python type checks |
pnpm typecheck:ts |
TypeScript only |
pnpm typecheck:py |
Pyright only |
pnpm test:frontend |
Full frontend Vitest suite |
pnpm validate:frontend |
TypeScript, frontend tests, and frontend build |
pnpm backend:test |
Backend pytest suite |
pnpm build:frontend |
Build renderer and Electron bundles |
pnpm setup:dev:win |
Windows development setup |
pnpm setup:dev:linux |
Linux development setup |
scripts/install-wangp-stack.ps1 |
Install/refresh WanGP GPU stack |
App data uses the AiVS folder name.
- Windows:
%LOCALAPPDATA%\AiVS\ - Linux:
$XDG_DATA_HOME/AiVS/or~/.local/share/AiVS/ - macOS:
~/Library/Application Support/AiVS/
Generated outputs are stored under the app data output directory and copied into project asset folders when saved to projects.
AGENTS_PRD.md- product direction and guardrailsAGENTS.md- coding-agent conventionsdocs/PHASE0_AUDIT.md- fork audit and preservation mapdocs/PHASE4_DETAILS.md- curated model profile briefbackend/architecture.md- backend architecturebackend/WANGP_BACKEND.md- WanGP bridge configurationscripts/wangp-stacks.json- curated GPU stack config
AiVS is changing quickly. Keep changes small, preserve inherited systems where possible, and route normal generation through WanGP only.
Before adding a model, add or update a curated profile in the backend profile registry. Do not expose arbitrary WanGP models directly in the UI.
Apache-2.0. See LICENSE.txt.
Third-party notices and model terms may apply to downloaded models and WanGP dependencies.