How to run Wan 2.1 locally.

1.3B or 14B.

Alibaba’s Wan team released Wan 2.1 on 25 February 2025. It has since been followed by Wan 2.2, but its 1.3B model is still one of the lightest ways to make video locally.

• T2V 1.3B is text to video on small cards. Wan’s figure is 8.19 GB of VRAM, and five seconds of 480p took about four minutes on an RTX 4090 without any speed-ups. Keep it at 480p: Wan calls 720p on the 1.3B less stable.
• T2V 14B is the full text-to-video model, at 480p or 720p, with about ten times the parameters.
• I2V 14B turns a picture into video. It comes in a 480p and a 720p version and needs a CLIP vision file on top of the usual three.

There are more: first-and-last-frame (FLF2V), VACE for editing and control, and the Fun models. They use their own workflows. If you are starting fresh with 8 GB, Wan 2.2 5B is the other option to try.

Files you need.

All from Comfy-Org’s repackage. The text encoder and VAE are the same for every Wan 2.1 model, and the 14B models of Wan 2.2 use them too.

Text to video, 1.3B

• Model wan2.1_t2v_1.3B_fp16.safetensors ComfyUI/models/diffusion_models/
  2.8 GB Download
• Text encoder umt5_xxl_fp8_e4m3fn_scaled.safetensors ComfyUI/models/text_encoders/ or umt5_xxl_fp16.safetensors, 11.4 GB
  6.7 GB Download
• VAE wan_2.1_vae.safetensors ComfyUI/models/vae/
  0.3 GB Download

Text to video, 14B

• Model wan2.1_t2v_14B_fp8_scaled.safetensors ComfyUI/models/diffusion_models/ or wan2.1_t2v_14B_fp16.safetensors, 28.6 GB
  14.3 GB Download

Image to video, 14B

• Model wan2.1_i2v_480p_14B_fp8_scaled.safetensors ComfyUI/models/diffusion_models/ or the fp16 file, 32.8 GB. A 720p version exists too
  16.4 GB Download
• CLIP vision clip_vision_h.safetensors ComfyUI/models/clip_vision/
  1.3 GB Download

models/
├── clip_vision/
│   └── clip_vision_h.safetensors        (image to video only)
├── diffusion_models/
│   └── wan2.1_t2v_1.3B_fp16.safetensors
├── text_encoders/
│   └── umt5_xxl_fp8_e4m3fn_scaled.safetensors
└── vae/
    └── wan_2.1_vae.safetensors

For the 14B there are two fp8 files. Pick fp8_scaled over plain fp8_e4m3fn: users rank the quality fp16, then bf16, then fp8_scaled, then fp8_e4m3fn.

Smaller files: GGUF

GGUF builds load through ComfyUI-GGUF with Unet Loader (GGUF). The encoder has one too, city96’s UMT5 GGUF, at 6.0 GB for Q8_0.

Q8_0 is close to fp16. On the 1.3B, GGUF saves little next to the 2.8 GB fp16 file; it matters for the 14B.

What fits your computer.

480p, 33 to 81 frames. ComfyUI loads the encoder first and moves it aside before sampling, so the model file sets the limit.

• 6 GBTight

  1.3B with ComfyUI’s offloading. Wan’s own figure is 8.19 GB, so expect it to lean on system RAM. Keep frames low.

• 8 GB1.3B fits

  1.3B at 832 × 480. The 14B doesn’t fit without GGUF and a lot of system RAM.

• 12 GBOffloads

  1.3B comfortably. 14B text to video as GGUF Q4_K_M or Q5_K_M. 14B image to video was too much for one 12 GB owner.

• 16 GBOffloads

  14B fp8 with part of it in system RAM, or GGUF Q5_K_M.

• 24 GBFits

  14B fp8 text to video and image to video.

• Mac 24 GB+1.3B fits

  1.3B fp16 with the euler sampler. M3 Max and M4 Max owners ran the example workflow; its uni_pc sampler gave them distorted video.

• Mac 64 GB+Tight

  14B as fp16 (28.6 GB) or GGUF, with the fp16 encoder. fp8 files don’t run on Apple GPUs.

• 12 GB graphics cards
• RTX 3060 12 GB
• 16 GB cards
• 24 GB cards
• Mac with Apple Silicon

Set it up.

1. 1

   Update ComfyUI

   Wan 2.1 has been in ComfyUI since 27 February 2025, so any current version runs it. ComfyUI Desktop updates itself. In a manual install:

   Terminal, in the ComfyUI folder Copy

   git pull
   pip install -r requirements.txt

2. 2

   Download the files

   Model, text encoder and VAE from the lists above, plus clip_vision_h for image to video. Use Comfy-Org’s files: a similarly named model from elsewhere gave one user only noise.

3. 3

   Put them in their folders

   Model in diffusion_models, encoder in text_encoders, VAE in vae, CLIP vision in clip_vision. Restart ComfyUI.

4. 4

   Open the template

   In the template browser, pick Wan 2.1 Text to Video or Wan 2.1 Image to Video. The text template is set up for the 1.3B.

5. 5

   Check the loaders

   Load CLIP gets UMT5 with type wan. Load VAE gets wan_2.1_vae. For image to video, Load CLIP Vision gets clip_vision_h, and your picture goes into WanImageToVideo.

6. 6

   Write a prompt and run

   Describe the scene and the motion. Keep the template’s Chinese negative prompt: it is Wan’s standard one.

Settings that work.

Comfy’s template follows Wan’s advice for the 1.3B. For the 14B, Wan’s own script uses a lower shift and CFG, more steps and more frames.

1.3B

Steps

30Wan’s script: 50

CFG

6Wan’s advice for 1.3B

Sampler

uni_pceuler on a Mac

Scheduler

simple

Shift

8Wan: 8 to 12

Size

832 × 480

Frames

3381 for five seconds

FPS

16

14B text to video, Wan’s settings

Steps

50Comfy template: 30

CFG

5Comfy template: 6

Sampler

uni_pceuler on a Mac

Scheduler

simple

Shift

5Comfy template: 8

Size

1280 × 720or 832 × 480

Frames

815 s at 16 fps

FPS

16

For image to video, Wan’s script uses 40 steps, with shift 3 at 832 × 480. Comfy’s template runs 20 steps at CFG 6, shift 8, 512 × 512 and 33 frames. Frame counts go in steps of four plus one, like 33, 49 or 81.

When it goes wrong.

ComfyUI shows “reconnecting” or a token error at the text encoder

The wrong text encoder. comfyanonymous’s answer: use the UMT5 file from Comfy-Org’s repackage.

mat1 and mat2 shapes cannot be multiplied (512x768 and 4096x5120)

A CLIP-L style encoder went into a model that expects UMT5. Load umt5_xxl_fp8_e4m3fn_scaled with type wan.

Only noise comes out

A similarly named model from another site. The Comfy-Org files fixed it.

Mac: distorted or banded video

The uni_pc sampler on Apple GPUs. Switch to euler.

Trying to convert Float8_e4m3fn to the MPS backend

fp8 files on a Mac. Use fp16 files or GGUF.

Image to video runs out of memory on 12 GB

The 14B I2V is heavy. People pointed to Wan 2.1 Fun 1.3B InP, a small image-to-video model, or GGUF.

ComfyUI doesn’t see the GGUF files

Install ComfyUI-GGUF and swap Load Diffusion Model for Unet Loader (GGUF).

Questions.

Sources: Wan2.1 repository, Wan2.1 generate.py, ComfyUI Wan 2.1 tutorial, Comfy-Org Wan 2.1 repackage, encoder thread, Mac output thread, ComfyUI issue #7027, WanVideoWrapper issue #461, length and audio thread.