Run Wan 2.2 on a cloud GPU

The Wan 2.2 model that fits DeployGPU's smallest card is TI2V-5B. The official README says it runs on a 24 GB GPU at 720p and 24 fps, and its own table shows a 22.9 GB peak. The two A14B models are the ones that want 80 GB: the README says its single-GPU command needs at least 80 GB, and its table shows a 59.8 GB peak at 720P and 41.3 GB at 480P with offloading on.

Three Wan 2.2 models and their measured peaks

The Wan2.2 README, checked 1 October 2026, lists T2V-A14B for text-to-video, I2V-A14B for image-to-video and TI2V-5B for both, all under Apache 2.0. The A14B models use two experts of about 14B parameters each, one for the early denoising steps and one for the late steps, which makes 27B in total with 14B active per step. TI2V-5B is dense with a high-compression VAE and generates 720P at 24 fps, at 1280 by 704 pixels.

The README includes a table of total time and peak GPU memory per run. Its single-GPU settings are --offload_model True --convert_model_dtype for the 14B models and the same two flags plus --t5_cpu for the 5B. The peaks below are read from that table, and the headroom is the card's memory minus the peak, calculated.

The peak was the same on H20, A100 and H100 class cards, but the time was not. For one 720P text-to-video run on a single GPU, the table lists 1,041.5 seconds on H100 or H800 and 2,735.7 seconds on A100 or A800, a ratio of about 2.6 (calculated). The README also says that on a GPU with at least 80 GB you can drop the offload flags for the 5B to speed it up. Its own text says TI2V-5B makes a 5-second 720P video in under 9 minutes on a single consumer-grade GPU, so expect different times on other cards.

ComfyUI: the repackaged files

ComfyUI's Wan2.2 tutorial uses the Comfy-Org/Wan_2.2_ComfyUI_Repackaged repository, and it says the 5B version should fit well on 8 GB of VRAM with ComfyUI's native offloading. Its templates are named "Wan2.2 5B" and "Wan2.2 14B T2V" in the template library. The file sizes below come from that repository, checked 1 October 2026.

The tutorial gives no VRAM figure for the 14B workflows, so use the README's peaks for planning. The 5B fits any 48 GB card, and I would start the A14B workflows on an 80 GB card:

Models go under /workspace/runpod-slim/ComfyUI/models/ inside the template's container, in the diffusion_models, text_encoders and vae folders the tutorial names. Run sudo docker exec deploygpu-workload ls /workspace/runpod-slim/ComfyUI/models to confirm the layout before you download. The template installs no models, the ComfyUI models guide covers downloading, and the ComfyUI template page explains why the private link cannot be called from code.

The README's command line on a GPU VM

For scripted runs, use a plain GPU VM. The README needs torch 2.4.0 or newer, and the PyTorch install guide covers that part. These are the README's steps for the 5B, with the download command spelled the way current huggingface_hub releases do:

git clone https://github.com/Wan-Video/Wan2.2.git
cd Wan2.2
pip install -r requirements.txt
pip install -U huggingface_hub
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 "Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage"

The README gives the older huggingface-cli download spelling for the same step. For T2V-A14B, the download and the single-GPU command are:

hf download Wan-AI/Wan2.2-T2V-A14B --local-dir ./Wan2.2-T2V-A14B
python generate.py --task t2v-A14B --size 1280*720 --ckpt_dir ./Wan2.2-T2V-A14B --offload_model True --convert_model_dtype --prompt "Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage."

The README says to add --t5_cpu if you hit out-of-memory errors. It also says that if flash_attn fails to install, you should install the other requirements first and flash_attn last. The A14B repositories are 126.18 GB each on Hugging Face, because the two experts are 57.15 GB each, and TI2V-5B is 34.18 GB. Compare those with the Disk column of the offer, and remember that stopping a DeployGPU instance deletes its disk, so the download repeats on the next launch. The stop guide covers what to copy off first, and the 80 GB page and the RTX PRO 6000 page compare the cards.

The README also describes S2V-14B for speech-to-video and Animate-14B for character animation. The FLUX page covers image models on the same ComfyUI template.