Wan2GP Desktop Launcher

The easiest way to run Wan2GP (WanGP) — the open-source generative video/image/audio toolkit — on Windows. One installer. One click to launch. Zero Python/CUDA setup.

   

⬇ Download for Windows — Latest Release
Wan2GP-Desktop-Launcher-*-win-x64.exe · ≈ 90 MB · Windows 10 / 11
⚠️ Unsigned installer — "unknown publisher" warning is normal for open-source without a code-signing cert.

📖 Visual walkthrough — install, Auto-Tune, dashboard & launch modes on one page →

⚡ Looking for lighter & faster? Try the Tauri edition

We recommend Wan2GP Desktop Launcher Tauri — same launcher, same Wan2GP, same features, new shell. This Electron edition ships its own Chromium + Node.js runtime inside every install; the Tauri edition uses the WebView2 engine already built into Windows and a compiled Rust backend.

	This Electron edition	Tauri edition
Installer download	≈ 97 MB	≈ 3 MB (~30× smaller)
Installed app	≈ 300+ MB (Chromium + Node)	≈ 7 MB
Idle RAM (launcher shell)	~200–400 MB	~30–80 MB
Backend	JavaScript on Node	Compiled Rust (no GC pauses)
Updates	Full ~97 MB re-download	Small patch

What that means for generation: the launcher doesn't render video — but every MB of RAM and VRAM it doesn't waste stays available for models. Its Launcher GPU setting (Integrated / Disabled) can push the UI off your NVIDIA card entirely, freeing 1–5 GB VRAM for Wan2GP.

Switching is painless: your Wan2GP install, models, wgp_config.json and desktop-config.json carry over untouched — the Tauri build follows this launcher's data-dir pointer automatically, and can silently uninstall this Electron edition afterwards (data kept).

---

Why Wan2GP? Why this launcher?

WanGP by deepbeepmeep is a one-stop super-app for open-source generative models — video, image, audio and TTS — with a full browser UI, queue, galleries, LoRAs, finetunes and plugins. It runs on as little as 6 GB VRAM and supports old and new GPUs alike.

This launcher handles it for you:

• One-click install — detects GPU, shows plan, installs everything
• Auto, per-GPU kernels from WanGP's setup_config.json, re-synced on every update
• Isolated uv env, pinned deps, no PATH editing
• One-click updates in Dashboard / Manage → Updates
• Install Wan2GP and Models (checkpoints, LoRAs, outputs) on any drive/folder you choose
• Auto-Tune recommends VRAM/RAM profile and writes config for you

---

Highlights — What you get with WanGP

Through the launcher you get the full WanGP — same models, same UI, same plugins. Nothing stripped.

Run on more hardware

• 6 GB VRAM is enough for select models — up to 24 GB+ for max quality/speed.
• NVIDIA: GTX 10xx / 16xx, RTX 20xx / 30xx / 40xx / 50xx. AMD: RDNA 2 / 3 / 3.5 / 4. Apple Silicon (via upstream).
• Quantized checkpoints: int8, fp8, GGUF, NV FP4, Nunchaku — architecture-aware downloads.
• Full web UI: galleries, reusable settings/templates, mask editor, background remover, pose/depth/flow, diarization, upsampling (RIFE/FlashVSR/Lanczos/SeedVR2), MMAudio/SeedVC, 20+ community plugins, LoRAs, finetunes, generation queue, headless/API mode.

Upstream docs: WanGP README · Installation · Models

---

Key features — What the launcher adds

• 🚀 One-click install — detects your GPU, shows exactly what it will install (Git, Python 3.11, PyTorch + CUDA, attention kernels), then does it. Missing Git/Python/uv? One click installs silently — no PATH editing. Reads NVIDIA RTX 20/30/40/50, AMD, Apple Silicon and picks the matching PyTorch + CUDA/ROCm build before installing.
• 🎯 Always the right kernels — per-GPU wheel set from WanGP's own setup_config.json. Re-syncs on install and every update. No stale wheels when upstream bumps them. Isolated Python 3.11 uv env with pinned deps (prebuilt wheels for pygame etc.).
• 📂 Clean data layout — C:\Wan2GP (app) + C:\Wan2GP-Models (models) by default, out of roaming AppData. Both are pre-filled defaults — pick any drive/folder at install.
• 🖥️ Flexible launch — Desktop (in-app), Browser, or External Terminal; pop-out, zoom, browser picker.
• 🔄 Safe updates — manual-only, version-aware. Nothing downloads without your action. Dashboard + Manage → Updates (WanGP core + launcher).
• 📂 Paths migrate — move installs between drives from Dashboard → Paths, no freeze, no leftovers, cross-drive safe.
• 🛡️ Crash-proof UI — renderer-crash watchdog, GPU-off recovery, no blank-screen regressions.

⚡ CUDA 13 stack on modern RTX cards. Since v2.4.5, RTX 20/30/40/50 get PyTorch 2.10 + CUDA 13 — SageAttention 2.2 (RTX 30/40) / 1.0.6 (RTX 20), FlashAttention 2.8.3, SpargeAttention (30/40/50), LightX2V (RTX 50), Nunchaku INT4/FP4 + GGUF 1.0.13 + bitsandbytes 0.49.2 (NF4). GTX 10/16 stay on CUDA 12.8 (no R580 needed); every other NVIDIA card needs R580+ and is checked before install.

---

Download & Install

Upgrading from v2.8.x? v3.0 moved Wan2GP out of roaming AppData — see Migration & Troubleshooting below. Clean reinstall is recommended.

1. Download the *.exe from Releases (button at top).
2. Run it — pick install + models folders (or accept C:\Wan2GP / C:\Wan2GP-Models). The screen detects your GPU and lists exactly what it will install — all paths are editable.
3. Click Install (~5–20 min: clone → uv venv → PyTorch+CUDA → requirements → kernels → wgp_config.json).
4. Click Launch — Desktop (green, in-app) or Browser (amber).

No Python, no CUDA toolkit, no pip needed beforehand — the installer fetches them.

💡 Creates Launch Wan2GP.bat desktop shortcut to run without the launcher.

Launch modes

• Desktop (green) — Wan2GP inside the launcher, with back/forward/reload, zoom 25–200% and pop-out to a separate window.
• Browser (amber) — visible console + auto-opens your browser when ready.
• External Terminal (blue) — real Windows Terminal / cmd via generated script; in-app LED + Stop kills by PID, closing the window also stops it.
• No-GPU Chrome — launch Chrome with GPU disabled to free VRAM for generation.
• Browser picker — detects Chrome, Edge, Firefox, Brave, Opera, Vivaldi.

Where is everything? (v3.0+, defaults)

C:\Wan2GP\                      ← repo + launcher data (self-contained)
   ├─ wgp.py                    ← Wan2GP core
   ├─ env_uv\                   ← Python 3.11 venv (uv)
   ├─ wgp_config.json           ← settings (ckpts → C:\Wan2GP-Models\ckpts)
   ├─ desktop-config.json       ← launcher config
   └─ boot.log                  ← diagnostic

C:\Wan2GP-Models\               ← your large files (any drive you chose)
   ├─ ckpts\                    ← checkpoints
   ├─ loras\                    ← LoRAs
   └─ outputs\                  ← generated videos/images/audio

C:\Wan2GP / C:\Wan2GP-Models are pre-filled defaults — Browse to any drive/folder at install or later via Dashboard → Migrate to new location. The dashboard shows a MODELS banner if it detects old AppData checkpoints.

Screenshots

---

⚡ Auto-Tune — one click, right profile

Manage → Auto-Tune (or ⚡ on the dashboard) scans GPU/VRAM/RAM/kernels and recommends the optimal wgp_config.json settings. All three profile dropdowns (video/image/audio) stay editable before you Apply.

The Manage → Settings tab also holds the GitHub token field (lifts GitHub API rate limit for update checks — see Migration & Troubleshooting below) and the Desktop → Auto-update toggle (launch-time checks / silent downloads / install-on-quit).

How profiles work — WanGP's memory manager (mmgp) uses 7 profiles trading VRAM for speed. Auto-Tune picks one from your VRAM × RAM:

Settings written to wgp_config.json:

• video_profile / image_profile / audio_profile — profile 1, 2, 3, 3.5, 4, 4.5, 5
• transformer_quantization — Scaled Int8 (recommended), FP8, NVFP4, or None
• enable_int8_kernels — Enabled if Triton available (default, experimental, ~10% faster with INT8 checkpoints)
• vae_config — always Auto (runtime picks VAE tiling from real VRAM headroom)
• vram_safety_coefficient — 0.80 (≥12GB), 0.70 (<12GB), 0.60 (failsafe) — forwarded as --vram-safety-coefficient on every launch (Extra Launch Args win)

Failsafe — tick "Prefer failsafe (P5 — maximum compatibility)" to force P5 regardless of matrix, for hardware where the recommendation still crashes.

---

📊 Monitoring & control

• Dockable console — live server log, dock to bottom/left/top or float. Search, export, resize. Toggle Ctrl+` or topbar button.
• Topbar sparklines — CPU/GPU/RAM/VRAM mini real-time charts.
• Running LED & Stop — status light + one-click server stop.
• System tray — minimize to tray, auto-start with Windows, notifications on server ready/stop.
• Keyboard shortcuts — Ctrl+` terminal, F12 DevTools picker, Esc/Ctrl+W close webview.
• Maintenance — update WanGP or Desktop Launcher from Dashboard or Manage → Updates, upgrade, reinstall, switch envs, or uninstall-with-backup from the UI. Dashboard → Paths migrates installs between drives.
• Renderer-crash watchdog — auto-reloads UI (bounded, no loops), restores mode and self-heals the embedded view. Generation is never touched — server runs in its own process.

---

🔧 GPU kernels — what gets installed per GPU

WanGP is faster with vendor kernels than stock PyTorch. The launcher reads WanGP's own setup_config.json and shows exactly what it will install — and re-syncs on every update (no stale wheels when upstream bumps one).

Per-GPU set: RTX 20 → Sage 1.0.6 + Flash + Nunchaku + GGUF + bnb. RTX 30/40 → add Sparge + Sage 2.2.0. RTX 50 → add LightX2V. All get bitsandbytes. Versions track setup_config.json — next update installs new wheels automatically.

What the 1-click covers (vs manual guide):

PyTorch matrix: RTX 20/30/40/50 → Py 3.11.14 + PyTorch 2.10 + CUDA 13.0/13.1 · GTX 10xx → Py 3.10.9 + PyTorch 2.7.1 + CUDA 12.8. Avoids 2.8.0 (RAM leak) + 2.9.0 (VAE VRAM bug).

GTX 10/16 stay on legacy CUDA 12.8 (no R580). Modern RTX needs R580+ (checked before install). Upstream: INSTALLATION.md

---

Deepy — your offline agent

Configure without editing JSON: Settings → Deepy or the Dashboard card.

• Disabled — Deepy off; keeps local Prompt Enhancer (Florence 2 + Llama 3.2 3B / Joy 8B).
• Deepy Zero — local, no account/key. Qwen3.5 VL 4B (recommended) / 9B / Qwen3.8 VL 27B.
• Deepy Prime — remote LLM via OpenCode (free, local models), Claude Code (claude-agent-sdk==0.1.66 pinned bridge) or Codex (paid), or local Qwen3.8 VL 27B (needs the 27B model + GGUF 1.0.14; auto-sets 32k context + Summarize). Prime exposes WanGP's MCP tools.

Switching live-re-renders the selector; Apply writes a consistent wgp_config.json (with backup). Also editable inside WanGP: Configuration → Prompt Enhancer / Deepy.

Remote LLM engines (Deepy Prime):

• 🟢 OpenCode — free & local. Routes to Llama.cpp / LM Studio (zero cost, no API key) or cloud. Install via npm → Start server → http://127.0.0.1:4096.
• 💲 Claude Code — paid. Max/Pro subscription or Anthropic API key (pay-per-token).
• 💲 OpenAI Codex — paid. Own CLI + OpenAI account.

New to this? Start with OpenCode — the only zero-cost option.

---

🧩 Plugin Manager — Status Pro included

Manage → Plugins lists Wan2GP's catalog merged with your installed plugins/ folder (system vs community grouping), with search, Name/Latest/Author sort, and per-plugin enable checkboxes. From a git URL you can install (clone + requirements.txt + enable), per-plugin ↻ check/update, 🗑 uninstall, library refresh, and check-all-updates — all with console progress.

• Status Pro is a default plugin: installed automatically on fresh setup and kept enabled (locked checkbox), but still uninstallable — one click reinstalls it.
• ★ Favourites auto-install on fresh setup (stored in desktop-config.json → favoritePlugins).
• Update badges (⇪) persist across restarts.
• Changes apply on next Wan2GP launch.

✨ DLSS5 installer — optional NVIDIA upsamplers

Dashboard card (below Deepy) runs WanGP's own scripts/install_dlss5.ps1 (workers v1.1.2, ReShade 6.8.0, RenoDX 4.70, DLSSNR 310.8.SF-v2, DLSS 310.8.0, Frame Generation 310.7.0) into dlss5/ with a live per-component checklist — downloading → SHA-256 ✓ → installed — plus console progress.

• Strict consent: type I ACCEPT (third-party binaries are community-hosted, unsigned, proprietary — see docs/DLSS5.md).
• Force backs up + replaces conflicting files. Stop Wan2GP first.
• Needs Windows 11 + RTX 30+ (Neural Rendering) / RTX 40+ (Frame Generation) + HAGS.

---

📦 Migration (v3.0) & Troubleshooting

v3.0 folders moved — old AppData → C:\Wan2GP + C:\Wan2GP-Models (click to expand)

What	Before (v2.8.x)	Now (v3.0+)
Repo + venv + config	%APPDATA%\wan2gp-desktop\Wan2GP\Wan2GP	C:\Wan2GP (default, any drive you choose)
Checkpoints	<repo>\ckpts	C:\Wan2GP-Models\ckpts
LoRAs / Outputs	<repo>\loras / outputs	C:\Wan2GP-Models\loras / outputs

Roaming AppData was bad for 10–100s GB of models (sync/quota). New defaults are top-level, separate, and all Browse-editable at install or via Dashboard → Migrate to new location.

Preferred upgrade: Manage → Uninstall (keep models) → close launcher → run new .exe → fresh C:\Wan2GP → point ckpts at C:\Wan2GP-Models. Experimental in-app Migrate also exists (back up first, no guarantee). v3.0.0 auto-migrated, v3.0.1 shows a Migrate dialog instead (no auto-move).

Auto-update & GitHub token: launcher checks wan2gp-desktop/releases/latest + upstream Wan2GP commits (cached 5 min). Anonymous GitHub API is 60 req/h/IP — you'll see rate limited if you restart often. Fix: Manage → Settings → GitHub token (classic PAT, public_repo scope) → Save → Restart → 5000 req/h on your token. Token is stored locally in desktop-config.json, never shipped in the .exe.

Troubleshooting — Z-Image crash, blank window (click to expand)

Z-Image crash Input type (BFloat16) and bias type (Half) should be same — known upstream bug (VAE loaded fp16 but latents are bf16 for ZImageTurbo_quanto_bf16_int8). Launcher fixes it automatically since v2.2.4: forces Z-Image VAE to bf16 at bootstrap ([bootstrap] z-image VAE dtype fix APPLIED). No action needed. Permanent fix is upstream PR #2095.

Blank / black window after update (title bar only): update left a locked app.asar. Uninstall → reinstall latest .exe with launcher fully closed. Still blank? Check %LOCALAPPDATA%\Wan2GP Desktop Launcher\boot.log: ready-to-show without first-paint = presentation class (create empty %USERPROFILE%\.wan2gp-desktop-gpu-off to disable HW accel); did-fail-load = corrupt bundle → reinstall. v2.8.5+ releases handles before swap and force-commits first frame to prevent both.

No prerequisites needed — launcher installs Git/Python/uv/Miniconda for you. For manual prerequisite help: PREREQUISITES.md.

---

🔥 What's New

Full history: docs/changelog.md · Each version below links to its standalone notes.

• v3.3.0 — 🧩 Plugin Manager (Status Pro default, favourites auto-install, persistent update badges), ✨ DLSS5 installer card, Deepy local Qwen3.8 Prime, Int8 Kernels auto-tune, boot update check, scoped stop sweep.
• v3.2.0 — Topbar cleanup: duplicate refresh icon removed, dead </> nav buttons removed, popout removed, red stop button, reload next to stop, labeled ⊞ Console button, metrics/title overlap fixed.
• v3.1.6 — Env hint moved/grey + smaller Check Updates + Desktop Full/Quick (delta vs 93 MB full) + deduped launch check + Shift+local.
• v3.1.5 — Full 93 MB update + sha512 verify (no blockmap delta) — fixes This app can't run after 3.1.3→3.1.4 auto-update.
• v3.1.4 — Minimal Desktop (backgroundThrottling only), console mirrors Terminal, auto-check + 5h polling with green dot, autoUpdate on by default.
• v3.1.3 — Manage → Updates tab, Dashboard → Paths non-blocking migrate (cross-drive, no leftovers), H3 Desktop shim, atomic config writes, GH_TOKEN leak fix, async GPU profile.
• v3.1.2 / v3.1.1 — Fix OpenCode/Codex Install via npm + Start server on spaced Node paths (C:\Program Files\nodejs) via services/spawn-cmd.js.
• v3.0.9 — Drive-root guard, co-located UV_CACHE_DIR (no hardlink warning), Manage → General Purge/Remove uv cache.
• v3.0.0 — Breaking: self-contained layout. Default C:\Wan2GP + separate C:\Wan2GP-Models (both user-selectable). Fresh per-GPU kernels from setup_config.json.

Older — v2.8 / v2.6

• v2.8.7 — Blank-screen root cause fixed (nested .screen regression).
• v2.8.5 — In-app update no longer blanks the launcher.
• v2.8.2 — GPU-compositor first-present fix + watchdog.
• v2.8.1 — AMD installs match AMD guide (NF4 kernels).
• v2.6.0 — Install/update/uninstall freeze fixes; VAE on Auto.

See docs/changelog.md for full list.

---

🛠 Build from source

git clone https://github.com/GKartist75/wan2gp-desktop.git
cd wan2gp-desktop
npm install
npm start          # dev
npm run build:win  # Windows NSIS installer

For a release with latest.yml (auto-update): GH_TOKEN=*** ./scripts/release-win.sh 3.0.9 — tags, pushes, and uploads exe + latest.yml + blockmap.

---

Documentation

---

Credits & License

Wan2GP Desktop Launcher wraps Wan2GP by deepbeepmeep. Released under the same License.

Discord: WanGP Community · X: @deepbeepmeep · Site: wangp.ai