deepbeepmeep/WanGP

What it does

WanGP is a browser-based front end for running open-source generative models locally. It wraps video, image, speech, music and 3D pipelines with a shared UI, model downloads, a generation queue and post-processing tools. The pitch is hardware access: the README states select models can run with as little as 4 GB of VRAM, with support for GTX 10XX, RTX 20XX and newer Nvidia cards plus AMD RDNA 2–4 README lines 36–51.

Beyond single generations, it adds a local assistant called Deepy Prime that plans and runs jobs, and experimental interactive "worlds" that can be walked through with the keyboard README lines 26–35. Supported families include Wan 2.1/2.2-derived models, MiniMax H3, LTX-2.x, Hunyuan Video, Flux, Qwen Image, several TTS models and Hunyuan3D/TRELLIS-style 3D models; the README lists them per modality README lines 10–19.

Quick start

The project offers four guided install routes: the bundled bat/sh scripts, a Pinokio community installer, the third-party Wan2GP Desktop launcher, and the experimental official WanGP Manager for Windows README lines 244–254. The script route:

• Windows: run scripts\install.bat, then scripts\run.bat.
• Linux/macOS: run scripts/install.sh, then scripts/run.sh.

The installer can auto-select a stack or walk you through package manager, environment name and install mode; an args.txt in the scripts folder accepts launch flags such as --advanced --open-browser README lines 261–284.

For a manual RTX 20XX–50XX install, the README gives a Conda sequence: create a Python 3.11.14 environment, install PyTorch 2.10.0 with the CUDA 13 build, then pip install -r requirements.txt, followed by a Triton wheel matching your GPU and PyTorch README lines 325–334. GTX 10XX uses a separate Python 3.10.9 / PyTorch 2.7.1 set of commands README lines 348–358. Launch the app with python wgp.py and check the in-app Guides tab for recommended starting models. Debian-based systems can instead run ./run-docker-cuda-deb.sh to detect the GPU and build a CUDA image.

Batch jobs work without the UI: save a queue in the web interface, then run python wgp.py --process my_queue.zip README lines 425–433.

Constraints

• Model support is tied to GPU generation and PyTorch/Triton versions. RTX 20XX must stay on Triton 3.2 with legacy compatibility, while RTX 50XX should use PyTorch 2.10 / Triton 3.6 for optimized INT8 and NV FP4 paths README lines 327–334.
• SageAttention 2 requires Ampere or newer; GTX 10XX cards fall back to SDPA. Linux prebuilt wheels are distributed from a separate fork README lines 325–347.
• Large model files are downloaded automatically and matched to hardware. The --offline flag restricts work to already-downloaded files README lines 52–74.
• Optional features have their own requirements: Civitai downloads need an API key, and Deepy's web access is off by default.
• The license is marked NOASSERTION on GitHub, so review the repository's license before redistribution.
• WanGP Manager is described as experimental and Windows-only.

HotRepo's take

No ranking or star-growth records are available for this repository in the collected data, so this assessment rests on the project materials rather than trend history. What stands out is breadth against a single hardware target: one interface covers several model families across modality, with quantized checkpoints, architecture-aware downloads and low-VRAM profiles, which is a different tradeoff from running one dedicated pipeline per model. The tradeoff is setup complexity — Triton, SageAttention and PyTorch versions are GPU-specific, and the README keeps a separate GTX 10XX path and an experimental Windows manager. Deepy's app-building and world exploration features are labeled previews, so treat them as evolving rather than stable.