Universal AI agent skill for generating images, videos, audio, and music with 200+ local models.
Install once, generate anywhere. Works with MiMoCode, Claude Code, Codex, Cursor, Gemini CLI, Hermes Agent, and any harness that supports the Agent Skills specification.
| Capability | Models |
|---|---|
| Generate images | Flux 2 Dev, Qwen Image 2512, Z-Image, Krea 2, HiDream, Ideogram, Chroma |
| Generate videos | Wan 2.2, HunyuanVideo 1.5, LTX-2, Kandinsky 5, SkyReels, LongCat |
| Edit images | Flux Kontext, Qwen Edit, DreamOmni2, USO, UMO |
| Edit videos | Lucy Edit, Kiwi Edit, Chrono Edit, Vace, Ditto |
| Animate characters | Multitalk, Infinitetalk, Hunyuan Avatar, SCAIL, Steady Dancer, Ovi |
| Generate speech | IndexTTS2, Chatterbox, Qwen3 TTS, OmniVoice, KugelAudio |
| Create music | ACE-Step, Stable Audio 3, HeartMuLa |
- Hardware-aware — auto-detects GPU/VRAM and recommends models that actually fit
- Model-specific prompting — each model family has different conventions; the agent knows them all
- Auto-presets — Lightning LoRAs for Qwen, Turbo LoRAs for Flux, applied automatically
- OOM recovery — detects out-of-memory errors and falls back to smaller models or fewer steps
- Browser gallery — live image viewer with auto-refresh, history, and zoom
- Auto-install — if Wan2GP isn't found, offers to install it via Pinokio headlessly
npx skills add PrimeEcto/Wan2AI -gThat's it. The skill is now available to all supported AI agents on your system.
Open your AI agent and invoke the skill with /wangp followed by your request:
/wangp Generate an image of a red fox in a snowy forest
/wangp Make a video of a sunset over the ocean
/wangp Create a talking head video from this portrait
/wangp Generate a lo-fi hip hop beat
The /wangp prefix tells the agent to load the Wan2AI skill. In most harnesses (MiMoCode, Codex, Claude Code, Cursor, Gemini CLI, Hermes Agent), this activates the skill so the agent follows the correct workflow.
Some harnesses may use a different syntax:
- MiMoCode / Codex:
/wangp <prompt>or just mention "generate an image" (auto-triggers) - Claude Code:
/wangp <prompt>or the agent may auto-detect - Hermes Agent: The agent invokes the skill automatically when you ask to generate media
The agent handles everything: hardware detection, model selection, prompt adaptation, generation, and display.
Wan2AI uses Wan2GP as its generation engine. You have three options:
If Wan2GP isn't installed, the agent will detect this and offer to install it automatically via Pinokio's headless CLI:
"Wan2GP is not installed. I can install it automatically via Pinokio. Want me to proceed?"
If you say yes, it runs:
npm install -g pinokio
pinokio download https://github.com/6Morpheus6/wan2gp
pinokio run ~/.pinokio/api/wan2gp/install.jsThis handles all dependencies, GPU setup, CUDA/torch installation, and model downloads — zero manual configuration.
- Download Pinokio
- Search for Wan2GP and click Install
- Launch Wan2GP from Pinokio
Follow the Wan2GP README. Requires Python 3.10+, CUDA (NVIDIA), ROCm (AMD), or MPS (Apple Silicon).
wangp/
├── SKILL.md # Agent workflow + design principles
├── agents/openai.yaml # UI metadata
├── scripts/
│ ├── wangp.py # Python CLI wrapping Wan2GP's API
│ └── viewer/
│ ├── server.py # Gallery HTTP server with SSE
│ ├── gallery.html # Dark-themed gallery UI
│ └── start.sh # Start script
└── references/
├── model-catalog.md # 200+ models organized by task
├── prompting.md # Per-family prompt conventions
└── hardware.md # GPU profiles and VRAM recommendations
The Python CLI (wangp.py) wraps Wan2GP's existing in-process API — no extra servers, no MCP dependencies, no code duplication.
# Detect hardware, Wan2GP version, and all installations
python scripts/wangp.py detect
# List 200+ models with optional filters
python scripts/wangp.py list --family flux
python scripts/wangp.py list --available
# Get default settings for a model
python scripts/wangp.py defaults qwen_image_2512_20B
# Generate with preset LoRA and gallery output
python scripts/wangp.py generate \
--model qwen_image_2512_20B \
--prompt "a red fox in snow" \
--lora "https://huggingface.co/DeepBeepMeep/Qwen_image/resolve/main/loras_accelerators/Qwen-Image-2512-Lightning-4steps-V1.0-bf16.safetensors" \
--output-dir /tmp/wangp-gallery \
--show
# Upscale an existing image
python scripts/wangp.py upscale image.jpg --method lanczos2
# Check for Wan2GP updates
python scripts/wangp.py updateWan2AI includes a built-in image gallery that opens in your browser. Images appear automatically as they're generated — no manual refresh needed.
Agent: "Would you like a live gallery in your browser?"
User: "Yes"
Agent: *starts server, opens browser*
*all generations stream to gallery automatically*
Features:
- Live auto-refresh with toast notifications for new images
- Gallery history with clickable thumbnails
- Click to zoom, arrow keys to navigate
- Image dimensions and filename metadata
- Dark theme optimized for viewing generated images
- Keyboard shortcuts:
←→navigate,Spacezoom,Escclose
| Model | Params | Steps | Best For |
|---|---|---|---|
qwen_image_2512_20B + Lightning |
20B | 4 | Text rendering + realism |
krea2_turbo |
~12B | 8 | Aesthetic photography |
ideogram4 + Turbo |
9.3B | 12 | Typography, layouts |
z_image |
6B | 8 | Fast generation |
flux + Turbo Alpha |
12B | 10 | General purpose |
flux2_klein_9B |
9B | 4 | Fast distilled Flux 2 |
| Model | Params | Steps | Best For |
|---|---|---|---|
t2v_sf (Lightning) |
14B | 4 | Fast text-to-video |
t2v_2_2 |
14B | 30 | Best quality T2V |
ltx2_22B_distilled |
22B | 8 | Fast with audio |
hunyuan_1_5_t2v |
8B | 30 | Lower VRAM |
i2v_2_2 |
14B | 30 | Image-to-video |
See references/model-catalog.md for the complete 200+ model list.
The agent automatically applies acceleration LoRAs when you request specific models:
| Request | Preset Applied | Steps |
|---|---|---|
| Qwen Image (any) | qwen_image_2512_20B + Lightning 4-step LoRA |
4 |
| Qwen Image (OOM) | Fallback to Lightning 8-step LoRA | 8 |
| Flux 2 Dev 32B | Turbo LoRA from HuggingFace | 8 |
| Flux 1 (any) | Turbo Alpha LoRA | 10 |
| Krea 2 | krea2_turbo (distilled) |
8 |
| Ideogram | Turbo preset | 12 |
If a generation runs out of memory, the agent detects it and automatically retries with a smaller model or fewer steps.
| VRAM | What You Can Run |
|---|---|
| 24 GB+ | Everything at full quality |
| 16 GB | Most models with int8 quantization |
| 12 GB | Image models (Qwen Lightning, Z-Image, Krea 2 Turbo), 14B video with int8 |
| 8 GB | 1.3B video models, GGUF variants, small image models |
| 6 GB | GGUF Q4 only |
The agent auto-detects your hardware and picks models that fit.
Wan2AI works with any agent that supports the Agent Skills specification:
| Agent | Install Command |
|---|---|
| MiMoCode | npx skills add PrimeEcto/Wan2AI -g |
| Codex | npx skills add PrimeEcto/Wan2AI -g |
| Claude Code | npx skills add PrimeEcto/Wan2AI -g |
| Cursor | npx skills add PrimeEcto/Wan2AI -g |
| Gemini CLI | npx skills add PrimeEcto/Wan2AI -g |
| Hermes Agent | npx skills add PrimeEcto/Wan2AI -g |
| Cline | npx skills add PrimeEcto/Wan2AI -g |
| Roo Code | npx skills add PrimeEcto/Wan2AI -g |
| GitHub Copilot | npx skills add PrimeEcto/Wan2AI -g |
| OpenCode | npx skills add PrimeEcto/Wan2AI -g |
Use -g for global installation (available to all agents).
| Problem | Solution |
|---|---|
| "Wan2GP not found" | Let the agent install it via Pinokio, or set WAN2GP_ROOT env var |
| "Unknown model" | Run python scripts/wangp.py list to see all 200+ models |
| Out of memory | Agent auto-detects OOM and retries with smaller model. Or use z_image, flux2_klein_4B |
| Slow generation | Use distilled variants: t2v_sf, krea2_turbo, ltx2_22B_distilled |
| Gallery not opening | Check that port isn't blocked. Server runs on 127.0.0.1 by default |
| Update available | Agent detects and offers to run python scripts/wangp.py update |
MIT