Wan 2.2 is an open-weight video model family released under the Apache 2.0 license. For most home GPUs, the TI2V-5B variant is the realistic starting point, because the larger A14B models need far more memory. This guide lists the official files, folders and steps for ComfyUI, and flags where the official sources disagree.
We have not benchmarked Wan 2.2 ourselves. Every figure here is attributed to the official page it comes from.
The Wan 2.2 variants
The official Wan 2.2 repo lists the first release on July 28, 2025. It later added S2V-14B on August 26, 2025, and Animate-14B on September 19, 2025.
This guide covers the three core video variants. The repo describes them as follows.
| Variant | Architecture | Task | Resolution | Repo VRAM claim |
|---|---|---|---|---|
| T2V-A14B | Mixture of experts, 27B total, 14B active | Text-to-video | 480P and 720P | “At least 80GB VRAM” on a single GPU |
| I2V-A14B | Mixture of experts, 27B total, 14B active | Image-to-video | 480P and 720P | “At least 80GB VRAM” on a single GPU |
| TI2V-5B | Dense | Text and image to video | 720P (1280×704 or 704×1280) at 24 fps | “At least 24GB VRAM (e.g, RTX 4090 GPU)” |
The TI2V-5B resolution and frame rate come from its Hugging Face model card. S2V-14B and Animate-14B are separate releases and are not covered here.
License
The repo lists Wan 2.2 under Apache 2.0. That is a permissive license for the model weights and code.
A permissive license does not make every output legal. You remain responsible for what you create and share.
VRAM: the official sources disagree
The repo says TI2V-5B needs “at least 24GB VRAM (e.g, RTX 4090 GPU).” Per the same page, the A14B models need “at least 80GB VRAM” on a single GPU.
The ComfyUI Wan 2.2 docs are more optimistic about the small model. They say 5B “should fit well on 8GB vram with the ComfyUI native offloading.”
Our reading is that both can hold. The repo figure likely assumes the model sits fully in GPU memory, while offloading moves parts of it to system memory.
Offloading generally trades speed for memory, so a smaller card may run slower. We have not confirmed either figure on our own hardware.
The Hugging Face card says TI2V-5B makes “a 5-second 720P video in under 9 minutes” on one consumer GPU. Treat that as the publisher’s figure, not ours.
Which variant to pick
The choice comes down to your GPU memory and whether you start from an image. Here is how the official figures sort the options.
- TI2V-5B: the only variant with a consumer-GPU claim. It takes text, or text plus a starting image, and outputs 720P at 24 fps per its model card.
- I2V-A14B: image-to-video at 480P or 720P. The repo’s figure of at least 80GB on one GPU puts it beyond most home setups.
- T2V-A14B: text-to-video only, with the same 80GB figure. It does not take a starting image.
If you have not used ComfyUI before, start with TI2V-5B. It needs only three files, and the ComfyUI docs provide a template for it.
Before you start
- Check how much VRAM your GPU has, and compare it with the figures above.
- Make sure you have enough free disk space for the model files. The ComfyUI docs page links each download.
- Prepare a starting image you created yourself, of a fictional adult. Never use a photo of a real person.
Files and folders
File names below are as listed in the ComfyUI Wan 2.2 docs. Folders are relative to your ComfyUI install, and the docs page links each download.
TI2V-5B (recommended starting point)
| File | Folder |
|---|---|
wan2.2_ti2v_5B_fp16.safetensors |
models/diffusion_models/ |
umt5_xxl_fp8_e4m3fn_scaled.safetensors |
models/text_encoders/ |
wan2.2_vae.safetensors |
models/vae/ |
14B image-to-video
| File | Folder |
|---|---|
wan2.2_i2v_high_noise_14B_fp16 |
models/diffusion_models/ |
wan2.2_i2v_low_noise_14B_fp16 |
models/diffusion_models/ |
umt5_xxl_fp8_e4m3fn_scaled.safetensors |
models/text_encoders/ |
wan_2.1_vae.safetensors |
models/vae/ |
Note that the 14B workflow uses the Wan 2.1 VAE, not the Wan 2.2 one, per the docs. The 14B models also come as a pair: a high-noise file and a low-noise file.
14B text-to-video
The docs list wan2.2_t2v_high_noise_14B_fp8_scaled and wan2.2_t2v_low_noise_14B_fp8_scaled. Both go in models/diffusion_models/.
Setup steps for TI2V-5B
- Install ComfyUI, or update an existing install. An older build may lack the nodes that the Wan 2.2 template needs.
- Download the three 5B files listed above from the links on the ComfyUI Wan 2.2 docs page.
- Place each file in its folder. The model goes in
models/diffusion_models/, the text encoder inmodels/text_encoders/and the VAE inmodels/vae/. - Restart ComfyUI so it finds the new files.
- Load the official Wan 2.2 TI2V-5B template workflow linked from the ComfyUI Wan 2.2 docs.
- Check that each loader in the workflow points to the matching file name.
- For image-to-video, add your starting image to the workflow’s image input. Use only an image you made of a fictional adult.
- Set the resolution to 1280×704 for landscape or 704×1280 for portrait, at 24 fps. These are the 720P sizes the model card lists.
- Write your prompt and run the workflow.
For prompt structure, see our guide to camera-motion prompts. The Prompt Lab can draft text descriptions of fictional adults for you.
Estimating local cost
Local runs use no vendor credits, but they do use electricity. The cost-per-clip calculator can help you estimate that cost from your own run times and power price.
Our guide to cloud vs local cost per clip compares this with paid apps.
Check current plans on OurDream AI
Browser-based option if you do not have a 24 GB GPU; publishes per-clip dreamcoin costs.
Troubleshooting
- Out of memory on a 14B model. Switch to TI2V-5B. The repo’s own figure for A14B is at least 80GB on one GPU.
- Out of memory on 5B. Make sure ComfyUI’s native offloading is enabled and close other apps that use the GPU. Then try a lower resolution.
- Lower resolution, odd results. The model card lists 720P as the TI2V-5B target. Moving away from it may reduce quality, so return to 1280×704 or 704×1280 when memory allows.
- File names do not match the guide. The ComfyUI page is itself inconsistent for 14B. Its I2V download list names fp16 I2V files, while its setup steps name fp8 T2V files. Point each loader at the files you actually downloaded.
- Wrong VAE. Per the docs, 5B uses
wan2.2_vae.safetensorsand the 14B models usewan_2.1_vae.safetensors. - Missing nodes or a template that will not load. Update ComfyUI, restart it and load the template again.
Safety: the rules are yours
No content filter is documented on the official Wan 2.2 pages we checked. A local model runs on your hardware, so the rules about what it makes are yours to keep.
Never animate photos of real people. That includes partners, former partners, coworkers, classmates and public figures.
Intimate fakes of identifiable people are illegal in many places. Our guide to AI video and real-person likeness law covers the US, UK and EU.
Use only images you created of fictional adults. Our consent policy explains the line this site holds.
For other local options, see open-weight image-to-video models compared.
Sources and dates
Facts checked against the linked official pages on October 31, 2026. Model files and docs change often, so check the ComfyUI page before you download.