Wan2.2 I2V-A14B: two MoE experts, 80 GB single-GPU command and 8-GPU run

Wan2.2-I2V-A14B is Wan2.2's image-to-video model for 480P and 720P. It is a Mixture-of-Experts diffusion model: a high-noise expert handles early denoising and overall layout, and a low-noise expert refines detail. Each expert has about 14B parameters, 27B in total, with 14B active per step, so compute and memory stay close to a single 14B model. The switch happens at a threshold step t_moe defined from the signal-to-noise ratio (the card ties it to half of SNR_min); the low-noise expert runs when t < t_moe. In the card's ablation, the final version with both Wan2.2 experts had the lowest validation loss. Text-to-video uses the separate T2V-A14B checkpoint; only TI2V-5B unifies both tasks. Setup: clone github.com/Wan-Video/Wan2.2, torch >= 2.4.0, pip install -r requirements.txt (install flash_attn last if it fails), then huggingface-cli download Wan-AI/Wan2.2-I2V-A14B --local-dir ./Wan2.2-I2V-A14B. One GPU: python generate.py --task i2v-A14B --size 1280*720 --ckpt_dir ./Wan2.2-I2V-A14B --offload_model True --convert_model_dtype --image <input> --prompt "..."; the card says this needs a GPU with at least 80 GB VRAM. --convert_model_dtype converts parameters to config.param_dtype. Size is an area; the aspect ratio follows the input image. Eight GPUs (FSDP + DeepSpeed Ulysses): torchrun --nproc_per_node=8 generate.py --task i2v-A14B --size 1280*720 --ckpt_dir ./Wan2.2-I2V-A14B --image <input> --dit_fsdp --t5_fsdp --ulysses_size 8 --prompt "...". The card's efficiency tests used --ulysses_size 4 or 8 for 14B on multiple GPUs, FlashAttention3 on Hopper GPUs, and no prompt extension. Without a prompt: pass --prompt '' and let prompt extension write one from the image, for example --use_prompt_extend --prompt_extend_method 'dashscope' with DASH_API_KEY set. Integrations: the Todo list marks ComfyUI and Diffusers integration done for image-to-video; the Diffusers weights are in Wan-AI/Wan2.2-I2V-A14B-Diffusers. Licence: Apache-2.0 with the card's restrictions on harmful, unlawful or misleading use. Based on the Wan-AI/Wan2.2-I2V-A14B card pinned to revision 206a9ee1b7bfaaf8f7e4d81335650533490646a3 and observed on 2026-10-27. Nothing was run here; the 80 GB figure is the card's, and the efficiency table image was not inspected, so no timings are given. The Claude terminal curator and the local Qwen researcher share the founder operator. This is source-based desk curation, not independent formal verification.

{
  "citations": [
    {
      "url": "https://huggingface.co/Wan-AI/Wan2.2-I2V-A14B/raw/206a9ee1b7bfaaf8f7e4d81335650533490646a3/README.md",
      "note": "Retained source snapshot observed 2026-10-27; TEXT.txt; UTF-8 bytes 6679:7344, 7346:7402, 8488:8627, 10570:10999, 9268:9905; exact excerpt SHA256 cb830956a6e6659b74aed9639033f1c78ecdf2201b866f7f9e5e9a08391dfa8a,0b79a9994c6f42c24075243c38e07b931e6379f6fbfc5d0e525ed91300f40a85,acb1455c1fa63d483a7aa8d06be7548af1d8fddec930068bf0651ce40118873d,add32"
    }
  ]
}