Wan 2.2 Local Video Generation Requirements and Setup Guide
By Furkan Gözükara · Updated Nov 21, 2026
Wan 2.2 local video generation requirements depend on the model you run: the 5B TI2V model is the light option, while the 14B text-to-video and image-to-video models use a Mixture-of-Experts design with 27B total and 14B active parameters. With SwarmUI on a ComfyUI backend, FP8 or GGUF model files and block swapping, Wan runs even on weaker PCs, and the older Wan 2.1 1.3B model starts at about 3.5 GB of VRAM. On Windows you also need Python, Git, FFmpeg, CUDA, cuDNN and the C++ build tools.
Wan 2.2 models at a glance
Wan 2.2 comes in three main variants. The 14B models split denoising between a high-noise expert, which builds the structure, and a low-noise expert, which refines detail, so each of them needs two diffusion model files.
| Variant | Parameters | Task | Key files |
|---|---|---|---|
| TI2V-5B | 5B | Text and image to video | wan2.2_ti2v_5B_fp16, wan2.2_vae |
| T2V-A14B | 27B total, 14B active (MoE) | Text to video | high-noise and low-noise 14B fp8_scaled experts, umt5_xxl text encoder, wan_2.1_vae |
| I2V-A14B | 27B total, 14B active (MoE) | Image to video | high-noise and low-noise 14B fp8_scaled experts, umt5_xxl text encoder, wan_2.1_vae |
According to the Wan team, Wan 2.2 was trained on 65.6% more images and 83.2% more videos than Wan 2.1, and the release overview lists 5-second clips at 720p and 24 fps in its base configuration. You do not have to collect these files by hand: the SwarmUI installer and model downloader fetches every Wan 2.2 model in one click and is available to members on Patreon.
VRAM and GPU requirements
The most detailed measured minimums come from the SECourses Wan 2.1 Gradio app, and they are a useful floor when you plan Wan work on your own card:
| Model | Output size | Reported minimum VRAM |
|---|---|---|
| Wan 2.1 text to video 1.3B | 480x832 | about 3.5 GB |
| Wan 2.1 video to video 1.3B | 480x832 | about 3.5 GB |
| Wan 2.1 text to video 14B | 720x1280 | about 17 GB (works lower, but slower) |
| Wan 2.1 image to video 14B | 720x1280 | about 17 GB (works lower, but slower) |
For Wan 2.2 in SwarmUI, the precision of the files you pick matters as much as the GPU:
- FP8 scaled files are the default for the 14B experts. In the SwarmUI beginner's guide to AI video, SwarmUI developer mcmonkey fits the Wan 2.1 14B models on an RTX 4090 and uses fp8 files for the best performance.
- GGUF files, such as Q8, lower VRAM use further; the SECourses SwarmUI tutorials use them for low-VRAM cards.
- FP16 is the heaviest. In the RunPod tutorial, the Wan 2.1 FP16 model used about 57 GB of VRAM on a 96 GB RTX PRO 6000.
- Block swapping in ComfyUI keeps part of the model in system RAM when VRAM runs out.
System RAM matters too. Loading two diffusion models at once, for example an image model and a video model in one text-to-image-to-video run, can run out of RAM. The Wan 2.2 training tutorial fixes RAM crashes in ComfyUI with the --cache-none argument.
GPU support is broad. The SECourses libraries are compiled for RTX 1000 through RTX 5000 cards, the Wan 2.1 app added support for older GPUs such as the RTX 2000 series, and Wan2GP supports all GPUs, including the RTX 5000 series, on Torch 2.9.1 and CUDA 13.
Software you need on Windows
Every local route needs the same base tools. Wan2GP lists Python 3.11, FFmpeg, CUDA 13 with cuDNN 9.17, the C++ tools, MSVC and Git, plus a recently updated NVIDIA driver, because it installs pre-compiled CUDA 13 wheels.
- The quickest path is the 1-click AI requirements installer, shown in this Windows setup video. The requirements installer guide explains it.
- The manual path is the Windows AI requirements post with its step-by-step video.
- If a package fails to build, read how to fix CUDA and C++ build errors.
- Extract installers to a short path such as
C:/wan_gen, because long directory paths cause errors.
Three ways to run Wan locally
| Route | Best for | Notes |
|---|---|---|
| SwarmUI with ComfyUI backend | Wan 2.2 text and image to video with tested presets | Model downloader, preset import, RIFE interpolation, 8-step Lightning presets |
| Wan2GP | Wan 2.2 Animate, VACE, Multitalk and the 5B models | Optimized VRAM use and automatic model downloads; speed LoRAs are applied manually |
| Wan 2.1 Gradio app | Wan 2.1 on small GPUs | From 4 GB VRAM, batch folders, video extension |
The ComfyUI 1-click installer, which SwarmUI needs as its backend, the Wan2GP installer and the Wan 2.1 Gradio app are available to members on Patreon. If you only need image to video on a 6 GB card, FramePack is another option. For general setup, see the ComfyUI install guide and the SwarmUI install guide.
Setting up Wan 2.2 in SwarmUI
The Wan 2.2 full tutorial follows these steps:
- Install or update the ComfyUI backend, then install or update SwarmUI.
- Configure ComfyUI as the SwarmUI backend.
- Run the model downloader and download the Wan 2.2 models.
- Import the SwarmUI presets; the automatic preset import script is the easiest way.
- Choose an image-to-video or text-to-video preset. For image to video, set Init Image Creativity to 0.
- Set the frame count and RIFE, write a prompt and generate while you watch the logs.
The SwarmUI post recommends these ComfyUI backend arguments: --disable-pinned-memory --use-sage-attention --enable-triton-backend --disable-comfy-compiler. It also warns that Sage Attention can, rarely, reduce quality or give black output on some models.
For prompts, the Wan prompt guide explains the basic formula of subject, scene and motion. For image to video, describe only the motion, because the model already sees the image.
Speed: what to expect
- 8-step LoRAs: the fast 8-step presets update uses the lightx2v Wan2.2-Lightning LoRAs, and the tutorial compares 8 steps with 20 steps.
- Wan 2.1 with CausVid at 8 steps: in the CausVid SwarmUI tutorial, a clip took 90 seconds on an RTX 5090 and 170 seconds on an RTX 3090 Ti; text to video at 576x1008 ran at 8.4 s/it versus 18.2 s/it.
- Sage Attention: 37% faster than Flash Attention in the Wan 2.1 app at 20 steps, 81 frames and 1280x720 with TeaCache 0.15.
- TeaCache: 2x faster and more at a 15% threshold with 30 steps and CFG 6, per the SwarmUI TeaCache tutorial.
Two problems slow generation more than anything else. When VRAM spills into shared memory, GPU usage drops, so balance resolution, frame count and model size against your VRAM. In the Wan 2.1 FusionX tutorial, disabling Sage Attention for image-to-video models fixed low GPU usage.
Frames, FPS and clip length
Wan 2.1 generates at 16 FPS, so the frame count is seconds x 16 + 1: 49 frames for 3 seconds and 81 frames, the Wan default, for 5 seconds. Fewer frames and a lower resolution, such as 640x640 instead of the SwarmUI default of 960x960 for Wan 14B, generate much faster with only a small quality loss. RIFE then doubles the frame rate, for example to 32 FPS.
No suitable GPU?
All three routes have installers for RunPod and Massed Compute, and the SwarmUI and Wan 2.1 app installers also cover SimplePod. See RunPod vs Massed Compute vs SimplePod to choose a platform.
FAQ
What is the minimum VRAM for Wan video generation?
The SECourses Wan 2.1 app runs the 1.3B model from about 3.5 GB of VRAM and the 14B models at 720p from about 17 GB, and lower VRAM still works, only slower. For Wan 2.2, SwarmUI with GGUF files and block swapping is the low-VRAM route, and Wan2GP is built around optimized VRAM use.
Does Wan 2.2 run on RTX 5000 series GPUs?
Yes. The ComfyUI and Wan2GP installers include RTX 5000 support, and Wan2GP installs Torch 2.9.1 with CUDA 13, Flash Attention 2.8.3, Sage Attention 2.2.0 and xFormers 0.0.34. Keep your NVIDIA driver updated for CUDA 13.
Is Wan 2.2 better than Wan 2.1?
The Wan 2.2 vs Wan 2.1 video shows a large difference on the same prompt. Wan 2.2 adds the Mixture-of-Experts design and a much larger training set.
Can I train Wan 2.2 LoRAs on my own GPU?
Yes. The Wan 2.2 LoRA training tutorial shares configs based on over 64 trainings that work on GPUs with as little as 6 GB of VRAM. See how to train a LoRA on a low-VRAM GPU.