How do I run Wan 2.2 video generation on a cloud GPU?

Short answer: Deploy the ComfyUI template, download the Wan 2.2 files from the Comfy-Org repack into ComfyUI/models/diffusion_models, text_encoders and vae, then open the Wan 2.2 workflow from ComfyUI’s templates. Start with TI2V-5B (text-to-video and image-to-video in one model, about 18 GB of files), which the ComfyUI tutorial says “should fit well on 8GB vram with the ComfyUI native offloading”; move to the 14B T2V or I2V pair (two FP8 experts of 14.29 GB each) when you want the larger model (ComfyUI Wan 2.2 tutorial, read 2026-11-03). The Wan 2.2 models are released under Apache 2.0.

Which Wan 2.2 model?

Model What it does Files for ComfyUI (Hugging Face sizes, read 2026-11-03)
TI2V-5B text-to-video and image-to-video wan2.2_ti2v_5B_fp16 10.0 GB, wan2.2_vae 1.41 GB, umt5_xxl_fp8_e4m3fn_scaled 6.74 GB
T2V-A14B text-to-video wan2.2_t2v_high_noise_14B_fp8_scaled and ..._low_noise_14B_fp8_scaled, 14.29 GB each; wan_2.1_vae 0.25 GB; the same text encoder
I2V-A14B image-to-video wan2.2_i2v_high_noise_14B and ..._low_noise_14B, 14.29 GB each in FP8 (28.58 GB each in FP16); wan_2.1_vae; the same text encoder

The 14B models use two experts, one for the high-noise steps and one for the low-noise steps, so you download both. VRAM need depends on model, precision and workflow: resolution and the number of frames (length) change it a lot. The Wan2.2 README (read 2026-11-03) says its own generate.py runs TI2V-5B “on a GPU with at least 24GB VRAM (e.g, RTX 4090 GPU)” and the 14B T2V command “on a GPU with at least 80GB VRAM”, both with its offloading flags; ComfyUI’s FP8 files and offloading lower that. For a card comparison, see RTX 4090 vs RTX 4090 48G vs RTX 5090 for image generation.

Step 1: Deploy the ComfyUI template

On /deploy, choose the ComfyUI template; the page then shows only the GPUs that can run it. The template installs ComfyUI without model weights (/docs, read 2026-11-03), so you download Wan 2.2 yourself. Open a terminal in JupyterLab or over SSH:

COMFY=$(find / -maxdepth 4 -type d -name ComfyUI 2>/dev/null | head -n 1); echo "$COMFY"
nvidia-smi
df -h "$COMFY"      # the 14B FP8 set is about 36 GB, the 5B set about 18 GB

Step 2: Download the files

The Comfy-Org repack keeps files under split_files/<folder>/, so download into a staging folder and move them into place. Install huggingface_hub first, then turn on cloudgpu-net for the download (see how to speed up model downloads).

pip install -U "huggingface_hub"
cloudgpu-net on
export HF_XET_HIGH_PERFORMANCE=1
REPO=Comfy-Org/Wan-2.2          # the ComfyUI docs link Comfy-Org/Wan_2.2_ComfyUI_Repackaged, which now redirects here

TI2V-5B:

hf download $REPO \
  split_files/diffusion_models/wan2.2_ti2v_5B_fp16.safetensors \
  split_files/vae/wan2.2_vae.safetensors \
  split_files/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors \
  --local-dir ~/wan-dl

T2V-A14B (FP8):

hf download $REPO \
  split_files/diffusion_models/wan2.2_t2v_high_noise_14B_fp8_scaled.safetensors \
  split_files/diffusion_models/wan2.2_t2v_low_noise_14B_fp8_scaled.safetensors \
  split_files/vae/wan_2.1_vae.safetensors \
  split_files/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors \
  --local-dir ~/wan-dl

For I2V-A14B, replace t2v with i2v in the two diffusion model names. Then move the files into ComfyUI’s folders:

for d in diffusion_models vae text_encoders; do
  mkdir -p "$COMFY/models/$d"
  mv ~/wan-dl/split_files/$d/* "$COMFY/models/$d/" 2>/dev/null
done
cloudgpu-net off
ls -lh "$COMFY"/models/{diffusion_models,vae,text_encoders}

Step 3: Open the workflow and render

Open ComfyUI through the public HTTPS URL on the instance row, or through an SSH tunnel to port 8188 if you want it private (the public URL has no login; see the SSH port forwarding guide). In ComfyUI, open Workflow → Browse Templates → Video and pick the Wan 2.2 workflow that matches your files. If no Wan 2.2 templates appear, your ComfyUI is older than the Wan 2.2 support; update it (git pull in the ComfyUI folder, then pip install -r requirements.txt after cloudgpu-net off) and restart it.

Per the ComfyUI tutorial, the video size and frame count are set in Wan22ImageToVideoLatent (5B) or EmptyHunyuanLatentVideo (14B T2V). The Wan2.2 README says TI2V-5B “supports 720P video generation at 24 FPS”, with 720P meaning 1280*704 or 704*1280 for that model. Start with a short clip at a small size, watch memory, then scale up:

nvidia-smi --query-gpu=memory.used,memory.total --format=csv -l 2

If you run out of memory, lower the resolution or length first, then use the FP8 files instead of FP16.

Alternative: the official Wan2.2 script on a Raw GPU

Without ComfyUI, deploy Raw GPU, clone Wan-Video/Wan2.2, install its requirements, download the weights and run generate.py. The README’s single-GPU command for TI2V-5B is:

hf download Wan-AI/Wan2.2-TI2V-5B --local-dir ./Wan2.2-TI2V-5B
python generate.py --task ti2v-5B --size 1280*704 --ckpt_dir ./Wan2.2-TI2V-5B \
  --offload_model True --convert_model_dtype --t5_cpu \
  --prompt "<your prompt>"

The README uses the older huggingface-cli download name; hf download is the current command. On an RTX 5090, install a PyTorch build for CUDA 12.8 or newer first (RTX 5090 “no kernel image” fix).

If you only need clips and not your own GPU, CloudGPU’s model API also lists Wan video models: create a job with POST /v1/videos, poll GET /v1/videos/{id} and download the MP4 (API docs, read 2026-11-03).

Save your videos, then destroy

Outputs land in $COMFY/output. Data on the instance is deleted when you destroy it, so copy the MP4 files off first (rsync or scp). Billing is per minute, rounded up, with a 1-minute minimum; the instance is billed while the weights download, and billing stops when you destroy it.

Frequently asked questions

How much VRAM does Wan 2.2 need?

The ComfyUI tutorial says TI2V-5B “should fit well on 8GB vram with the ComfyUI native offloading”. It gives no figure for the 14B models; the Wan2.2 README’s own script asks for at least 24 GB for TI2V-5B and at least 80 GB for 14B T2V. VRAM need depends on model, precision and workflow, especially resolution and frame count (read 2026-11-03).

Which files does ComfyUI need for Wan 2.2 TI2V-5B?

wan2.2_ti2v_5B_fp16.safetensors in models/diffusion_models, wan2.2_vae.safetensors in models/vae, and umt5_xxl_fp8_e4m3fn_scaled.safetensors in models/text_encoders, from the Comfy-Org Wan 2.2 repack on Hugging Face.

Why are there two 14B model files?

Wan 2.2 A14B uses a high-noise expert and a low-noise expert, and the ComfyUI workflow loads both, so download both files for T2V or I2V.

Can I use Wan 2.2 output commercially?

The ComfyUI tutorial says the models “are based on the Apache 2.0 open source license and support commercial use”, and the Wan2.2 README says the models are licensed under Apache 2.0. Check the licence on the model page for the files you use.

What frame rate does Wan 2.2 produce?

The Wan2.2 README says TI2V-5B supports 720P video generation at 24 FPS.

Sources (read 2026-11-03, Asia/Shanghai)

The Deploy page lists the cards you can start. Prices are on https://cloudgpu.app/pricing and are not repeated here.

Last updated: 2026-11-03 (Asia/Shanghai).