deepbeepmeep/Wan2GP

How to use it

Install is conda + pip; the page gives exact versions. Newer GPUs (RTX 30XX+):

git clone https://github.com/deepbeepmeep/Wan2GP.git
cd Wan2GP
conda create -n wan2gp python=3.11.14
conda activate wan2gp
pip install torch==2.10.0 torchvision==0.25.0 torchaudio==2.10.0 --index-url https://download.pytorch.org/whl/cu130
pip install -r requirements.txt

An older-GPU path uses Python 3.10.9, torch 2.7.1, CUDA 12.8. Launch with python wgp.py. There are also 1-click .bat/.sh installers that additionally fetch acceleration kernels (Triton, Sage, Flash, GGUF, Lightx2v, Nunchaku), a Debian Docker script (./run-docker-cuda-deb.sh, ships CUDA 12.4.1 / PyTorch 2.6.0), and a Pinokio installer (the page recommends the community wan2gp / wan2gp-amd scripts over the official Pinokio one).

Headless mode processes a queue you build and save in the UI, no web server:

python wgp.py --process my_queue.zip

Hardware specifics the summary omits

The "6GB VRAM" headline is per-model, not universal. Concrete figures the page states: MiniMax H3 needs 5–6GB for 5s (124 frames) at 832x480, 8–9GB for 15s. Bernini 14B video needs 12GB (v2v) / 16GB (v2v + reference frames). SenseNova-U1.5 does native 4K in 8–9GB.

Acceleration is tightly version-gated:

• Triton: 3.3.x with PyTorch 2.7; 3.6.x with PyTorch 2.10 on RTX 30XX+; RTX 20XX must stay on 3.2 (and cannot use the optimized INT8/NVFP4 paths). Triton 3.2 on RTX 30XX+ crashes YuE2's INT8 ConvRot kernels.
• SageAttention: 1.0.6 on RTX 20XX, 2.2.0 on RTX 30XX+ (Ampere-or-newer); GTX 10XX must use SDPA.
• Sol-Attn (H3 sparse attention) requires BF16, Triton 3.6+, and RTX 40/50, H100/H200, or B100/B200.

Deepy agent: real requirements

The agentic assistant has two tiers. Deepy Zero is lightweight; Deepy Prime (Qwen3.8 VL 27B) does multi-step planning but requires 24GB VRAM — with a documented 16GB workaround (Qwen3.8 VL 27B GGUF Q2, 32000 KV cache, speculative decoding off, browser in no-GPU mode). Prime can also be driven by a remote Anthropic Claude or OpenAI Codex account (including some free accounts) instead of local weights, configured under Configuration / Prompt Enhancer / Deepy. With filesystem access set to read/write, it runs batch requests like "make one video for every prompt in this text file," scoped to WanGP outputs and folders you select.

Limits

Free to use locally; the page points to the license for terms but does not name it here. Deepy Prime's 24GB floor is a hard gate for most consumer cards without the degraded 16GB path. Some features are flagged experimental by the author (Scail 2 Parallel Subwindows — "may end up a big fail and removed"). Ideogram v4 has a hardcoded safety filter that can trip on harmless prompts with no workaround. This is a fast-moving single-maintainer project (changelog runs from v12.x through v13.13xx), so pinning a version matters.