The Open Generative AI desktop app acts as a thin client for Wan2GP: you run the Wan2GP Gradio server yourself on a machine with a capable GPU, and the desktop app sends prompts to it and retrieves the results over HTTP. The app bundles no Python environment, no model weights, and no GPU runtime for this engine — all inference happens on your server. This split design is intentional. Wan2GP’s runtime stack (Sage attention, flash-attn, AWQ/GGUF quantization kernels) is CUDA-only and has no Apple Silicon MPS path. By treating Wan2GP as a remote service, a Mac user can keep the familiar desktop UI while routing the heavy compute to a Linux or Windows GPU box, a gaming PC on the local network, or a rented GPU cloud instance.Documentation Index
Fetch the complete documentation index at: https://mintlify.com/anil-matcha/open-generative-ai/llms.txt
Use this file to discover all available pages before exploring further.
Why a Separate Server?
Wan2GP’s dependencies — Sage attention, flash-attn, and the AWQ/GGUF kernels used for memory-efficient inference — require CUDA. There is no working MPS (Apple Silicon) path in Wan2GP’s runtime. Rather than blocking Mac users from accessing video models entirely, the app exposes Wan2GP as a configurable remote endpoint. This means:- Mac users can keep the desktop app on their MacBook and offload generation to a Linux/Windows GPU machine on the same LAN or to a rented GPU instance.
- Linux/Windows users can run both the app and the Wan2GP server on the same machine, connecting to
http://localhost:7860. - Remote GPU users can spin up a RunPod or vast.ai instance, start Wan2GP there, and point the desktop app at the public URL.
Setting Up the Wan2GP Server
Clone the Wan2GP repository on your GPU machine
Run this on the Linux or Windows machine that has the NVIDIA or AMD GPU:
Run the installer
install.bat instead. The installer sets up the Python virtual environment, installs PyTorch, and downloads the required attention kernels.Start the server bound to all interfaces
--server-name 0.0.0.0 flag binds to all network interfaces so the desktop app can reach the server over LAN or a public IP. By default Wan2GP listens on port 7860. Note the IP address of this machine — you will need it in the next step.Connect the desktop app
In the Open Generative AI desktop app:
- Open Settings → Local Models
- Find the Wan2GP server section
- Paste the server URL — for example
http://192.168.1.42:7860for a LAN machine, orhttp://localhost:7860if running on the same machine - Click Test to verify the connection
- Click Save
Available Models
| Model | Type | Notes |
|---|---|---|
| Flux.1 Dev | Image | 1024px output, 28 steps. High-quality text-to-image. |
| Qwen Image | Image | 1024px output, 30 steps. Strong prompt adherence. |
| Wan 2.2 (T2V / I2V) | Video | Text-to-video and image-to-video. Slow on consumer GPUs. |
| Hunyuan Video | Video | High-quality text-to-video. One of the best open-source video models. |
| LTX Video | Video | Fastest video generation option available via Wan2GP. |
Image models in Image Studio — Flux.1 Dev and Qwen Image appear in Image Studio once a Wan2GP server is connected. Full Video Studio wiring for the video models is on the roadmap; video generation via Wan2GP is available through the generation API in the current release.
Supported GPU Platforms
| Platform | Supported |
|---|---|
| NVIDIA (CUDA) | ✅ Yes |
| AMD (ROCm) | ✅ Yes |
| Apple Silicon (MPS) | ❌ No |
| CPU-only | ❌ No |
Remote Deployment Options
If you do not have a local GPU machine, you can rent one and run Wan2GP there:- RunPod — Rent a GPU pod, clone Wan2GP, start the server with
--server-name 0.0.0.0, and use the pod’s public IP and port in the desktop app’s Wan2GP URL setting. - vast.ai — Similar workflow. Rent an instance with a CUDA GPU, expose port 7860, and point the desktop app at the instance’s public address.
- Local LAN gaming PC — If you have a Windows PC with an NVIDIA GPU on your home network, install Wan2GP there (
install.bat) and point the desktop app at the PC’s LAN IP.