Step 1: Use the Built-In Workflow Templates
ComfyUI ships Wan 2.2 workflows in its template library — no manual node wiring required. Open ComfyUI and go to Workflow → Browse Templates → Video, then load 'Wan2.2 14B T2V' for text-to-video, 'Wan2.2 14B I2V' for image-to-video, or 'Wan2.2 5B' for the lightweight hybrid model.
The 14B models use a two-expert MoE architecture (high-noise and low-noise experts) that ComfyUI loads sequentially, keeping peak VRAM under 24GB.
Step 2: Download the Model Files
All official files live in the Comfy-Org/Wan_2.2_ComfyUI_Repackaged repository on Hugging Face. You need four file groups:
- Diffusion models — wan2.2_t2v_high_noise_14B_fp8_scaled.safetensors + low_noise variant (FP8, ~14.3GB each)
- Text encoder — umt5_xxl_fp8_e4m3fn_scaled.safetensors
- VAE — wan_2.1_vae.safetensors (shared across the Wan 2.x family)
- Optional LoRAs — community relighting, style and motion LoRAs
Step 3: Place Files in the Right Folders
Drop diffusion models into ComfyUI/models/diffusion_models/, the text encoder into ComfyUI/models/text_encoders/, and the VAE into ComfyUI/models/vae/. Restart ComfyUI and the workflow templates will auto-locate them.
Step 4: Recommended Settings
For a first render on a 24GB card: 1280×720 resolution, 81 frames (~5 seconds at 16fps), 30 steps, CFG 5.0. If you hit out-of-memory at the VAE decode stage, drop to 1280×704 or 960×544 before touching anything else.
Low VRAM? Use GGUF Quantized Weights
Community GGUF quants (Q5_K_M ≈ 10.8GB, Q6_K ≈ 12GB) run the 14B experts on 12–16GB cards via the ComfyUI-GGUF custom node. Expect a modest quality trade-off in exchange for fitting in memory.
When to Skip Local and Use the API
Local is free per-video but slow on consumer hardware — expect several minutes per clip plus fiddling time. If you ship content daily, the hosted generator (and the Wan API) is faster per result, needs no VRAM, and always runs the newest model versions. Try the online generator free, or read our VRAM requirements guide if you're still deciding on hardware.