How to run Wan 2.2 locally.

5B or 14B.

Alibaba’s Wan team released Wan 2.2 on 28 July 2025, with ComfyUI support the same day. Three models are made for this kind of use, and they are built differently.

• TI2V 5B is one dense 5B model that does text to video and image to video. It makes 720p at 24 fps, which for this model means 1280 × 704, and 121 frames, five seconds. It has its own VAE.
• T2V A14B is text to video from two 14B experts. The high-noise model does the early steps, where layout and motion are set, and the low-noise model does the late steps, where detail goes in. You need both files, and ComfyUI runs them one after the other.
• I2V A14B is the same pair for image to video. Unlike Wan 2.1’s image model, it needs no CLIP vision file.

Pick 5B for an 8 or 12 GB card, a Mac, or when you want quick tries at 720p. Pick 14B when quality comes first and you have 16 GB or more, or 12 GB with GGUF files and patience. One RTX 3060 owner who moved from 5B to 14B GGUF found the quality much better. Wan 2.2 also has speech-to-video and Animate models; they need their own workflows and aren’t covered here.

Files you need.

Every version takes the same text encoder, UMT5-XXL. The VAE is where people go wrong: only the 5B uses the Wan 2.2 VAE. Both 14B models use the Wan 2.1 VAE.

Wan 2.2 5B

• Model wan2.2_ti2v_5B_fp16.safetensors ComfyUI/models/diffusion_models/
  10.0 GB Download
• Text encoder umt5_xxl_fp8_e4m3fn_scaled.safetensors ComfyUI/models/text_encoders/
  6.7 GB Download
• VAE wan2.2_vae.safetensors ComfyUI/models/vae/
  1.4 GB Download

Wan 2.2 14B, text to video

• High-noise model wan2.2_t2v_high_noise_14B_fp8_scaled.safetensors ComfyUI/models/diffusion_models/ or the fp16 file, 28.6 GB
  14.3 GB Download
• Low-noise model wan2.2_t2v_low_noise_14B_fp8_scaled.safetensors ComfyUI/models/diffusion_models/ or the fp16 file, 28.6 GB
  14.3 GB Download
• Text encoder umt5_xxl_fp8_e4m3fn_scaled.safetensors ComfyUI/models/text_encoders/
  6.7 GB Download
• VAE wan_2.1_vae.safetensors ComfyUI/models/vae/
  0.3 GB Download
• 4-step LoRA, high wan2.2_t2v_lightx2v_4steps_lora_v1.1_high_noise.safetensors ComfyUI/models/loras/ optional, the Comfy template uses it
  1.2 GB Download
• 4-step LoRA, low wan2.2_t2v_lightx2v_4steps_lora_v1.1_low_noise.safetensors ComfyUI/models/loras/ optional, the Comfy template uses it
  1.2 GB Download

Wan 2.2 14B, image to video

The same encoder and the same Wan 2.1 VAE, with the I2V pair and its own LoRAs. T2V LoRAs damage I2V output, so keep them apart.

• High-noise model wan2.2_i2v_high_noise_14B_fp8_scaled.safetensors ComfyUI/models/diffusion_models/ or the fp16 file, 28.6 GB
  14.3 GB Download
• Low-noise model wan2.2_i2v_low_noise_14B_fp8_scaled.safetensors ComfyUI/models/diffusion_models/ or the fp16 file, 28.6 GB
  14.3 GB Download
• 4-step LoRA, high wan2.2_i2v_lightx2v_4steps_lora_v1_high_noise.safetensors ComfyUI/models/loras/ optional
  1.2 GB Download
• 4-step LoRA, low wan2.2_i2v_lightx2v_4steps_lora_v1_low_noise.safetensors ComfyUI/models/loras/ optional
  1.2 GB Download

models/
├── diffusion_models/
│   ├── wan2.2_t2v_high_noise_14B_fp8_scaled.safetensors
│   └── wan2.2_t2v_low_noise_14B_fp8_scaled.safetensors
├── loras/
│   ├── wan2.2_t2v_lightx2v_4steps_lora_v1.1_high_noise.safetensors
│   └── wan2.2_t2v_lightx2v_4steps_lora_v1.1_low_noise.safetensors
├── text_encoders/
│   └── umt5_xxl_fp8_e4m3fn_scaled.safetensors
└── vae/
    └── wan_2.1_vae.safetensors

fp16 files look a little better than fp8 and take twice the space. Among fp8 files, fp8_scaled beats plain fp8_e4m3fn. Comfy-Org has no fp8 build of the 5B.

Smaller files: GGUF

QuantStack’s GGUF builds load through the ComfyUI-GGUF nodes with Unet Loader (GGUF). For 14B the sizes below are per half: download the HighNoise and the LowNoise file of the same level.

14B sizes are per half, so double them. Q8_0 is close to fp16. Q4_K_M and Q5_K_M are what most 12 and 16 GB owners run. The 14B T2V also has a Q2_K at 5.3 GB.

What fits your computer.

Video needs far more working memory than an image, and it grows with size and frame count. ComfyUI moves what doesn’t fit into system RAM, so most setups below still run, only slower.

• 6 GBTight

  5B as a small GGUF (Q4_K_M is 3.4 GB) at a smaller size with fewer frames. No reliable report at this tier; 8 GB is the practical floor.

• 8 GB5B fits

  ComfyUI’s own docs say the 5B fits 8 GB with its offloading. For 14B, the advice given on Hugging Face is GGUF Q2 or Q3 per half.

• 12 GBOffloads

  5B comfortably. 14B as Q4 GGUF: an RTX 3060 with 32 GB of RAM makes five seconds in 10 to 12 minutes with a 4-step LoRA.

• 16 GBOffloads

  14B as GGUF Q4 to Q6, or fp8 with part of it in RAM. A 16 GB RTX 4090 laptop ran I2V Q3_K_M without spilling into RAM.

• 24 GBFits

  14B fp8 pair at 640 × 640, 81 frames: Comfy measured about 84% of an RTX 4090D’s memory.

• Mac 16 to 24 GBTight

  5B as fp16 or GGUF, with the euler sampler. fp8 files don’t run on Apple GPUs.

• Mac 32 to 48 GB5B fits

  5B at 832 × 480, 121 frames ran on a 48 GB M4 Max at about 39 s per step. 14B only as GGUF.

• Mac 64 GB+Tight

  The 14B fp16 pair is 57.2 GB of weights, so 96 to 128 GB is more realistic, or GGUF. A 128 GB M1 Ultra gave black video at 81 frames.

• 12 GB graphics cards
• RTX 3060 12 GB
• 16 GB cards
• RTX 5060 Ti 16 GB
• 24 GB cards
• RTX 5090
• Mac with Apple Silicon

Set it up.

1. 1

   Update ComfyUI

   An old ComfyUI is behind the most common Wan 2.2 error, the 36-channel one below. ComfyUI Desktop updates itself; the portable build has an update script, and if that doesn’t take, a fresh portable copy does. In a manual install:

   Terminal, in the ComfyUI folder Copy

   git pull
   pip install -r requirements.txt

2. 2

   Download the files

   From the lists above: model (or both halves), text encoder and the right VAE. Add the two 4-step LoRAs if you want the fast template.

3. 3

   Put them in their folders

   Models in diffusion_models, encoder in text_encoders, VAE in vae, LoRAs in loras. Restart ComfyUI so the loaders list them.

4. 4

   Open the template

   In the template browser, pick Wan 2.2 5B Video Generation, Wan 2.2 14B Text to Video or Wan 2.2 14B Image to Video. The 14B templates come with the 4-step LoRAs switched on.

5. 5

   Check the loaders

   Load CLIP gets UMT5 with type wan. Load VAE gets wan2.2_vae for 5B and wan_2.1_vae for 14B. On 14B there are two Load Diffusion Model nodes: high-noise feeds the first KSamplerAdvanced, low-noise the second. For image to video, the 5B takes your picture in Wan22ImageToVideoLatent, the 14B in WanImageToVideo.

6. 6

   Write a prompt and run

   Describe the scene and the motion. The templates carry Wan’s standard Chinese negative prompt; leave it in. The first run loads everything and takes longest.

Settings that work.

Wan’s own settings and Comfy’s templates differ. Wan’s scripts use more steps and a different shift; Comfy trades some quality for speed. Both are listed so you can choose.

5B, Wan’s settings

Steps

50Comfy template: 20

CFG

5

Sampler

uni_pceuler on a Mac

Scheduler

simple

Shift

5Comfy template: 8

Size

1280 × 704or 704 × 1280

Frames

1215 s at 24 fps

Text encoder

Load CLIP, wan

14B text to video, Wan’s settings

Steps

40Comfy examples: 20

CFG

4 / 3high / low. Comfy: 3.5

Sampler

euleras in Comfy’s examples

Shift

12Comfy examples: 8

Size

1280 × 720or 832 × 480

Frames

815 s at 16 fps

Switch

by noise levelComfy: half the steps each

Text encoder

Load CLIP, wan

For 14B image to video, Wan uses 40 steps, CFG 3.5 on both halves and shift 5. Comfy’s older non-LoRA example uses 20 steps, CFG 3.5 and shift 8. In KSamplerAdvanced, the high-noise sampler runs from step 0 to half, with leftover noise returned, and the low-noise sampler finishes.

14B with the 4-step LoRAs (Comfy’s template)

Steps

42 high, then 2 low

CFG

1

Sampler

euler

Scheduler

simple

Shift

5on both halves

Size

640 × 640

Frames

81at 16 fps

LoRA strength

1.0one LoRA per half

Four steps is the total, not per model. The LoRAs make motion slower, which lightx2v acknowledges. What helped people: a larger size (1280 × 720 moved at normal speed where 832 × 480 was slow), a stronger high-noise LoRA, newer LoRA releases, or a third sampler that runs one or two high-noise steps without the LoRA at CFG 3.5 first.

How fast.

The first time includes loading the models. On the RTX 3060 with a LoRA, text to video took 10 to 12 minutes and image to video about 15. Wan’s own figure for the 5B is under nine minutes for five seconds of 720p on one consumer GPU.

When it goes wrong.

weight of size [5120, 36, 1, 2, 2], expected input[1, 32, 21, 96, 96] to have 36 channels, but got 32 channels instead

The classic 14B image-to-video error. Update ComfyUI (a fresh portable copy if the updater doesn’t work). Delete the WanImageToVideo (Flow2) node pack, which overwrites core code. And use wan_2.1_vae, not the Wan 2.2 VAE.

expected input[1, 16, …] to have 48 channels, but got 16 channels instead

The Wan 2.2 VAE on a 14B model. 14B uses wan_2.1_vae.safetensors; only the 5B takes wan2.2_vae.

size mismatch for encoder.conv1.weight

Wrong VAE for the model. Same fix: the 2.1 VAE for 14B, the 2.2 VAE for 5B.

mat1 and mat2 shapes cannot be multiplied (462x768 and 4096x5120)

The wrong text encoder. Use umt5_xxl_fp8_e4m3fn_scaled from Comfy-Org in Load CLIP with type wan.

lora key not loaded: blocks.0.cross_attn…

Some original lightx2v LoRA files don’t match ComfyUI’s key names. Kijai’s converted versions in Kijai/WanVideo_comfy load cleanly.

Everything moves in slow motion

A side effect of the 4-step LoRAs. Go larger, raise the high-noise LoRA strength, or add a first high-noise pass without the LoRA, as described above.

The 5B makes only brown fog

A broken install. A fresh ComfyUI with the models moved over fixed it.

Mac: the first frame is fine, then mush or bands

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.

ComfyUI gets killed or freezes on a 32 GB PC

Out of system RAM with the fp8 pair. Use GGUF, add swap, or try --cache-ram.

ComfyUI doesn’t see the GGUF files

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

Questions.

Sources: Wan2.2 repository, TI2V-5B model card, T2V-A14B config, ComfyUI Wan 2.2 tutorial, Comfy-Org workflow templates [1], ComfyUI examples, QuantStack GGUF thread [2], Civitai RTX 3060 workflow [3], ComfyUI issue #16644 [4], 36-channel thread, ComfyUI issue #9092, uni_pc on Mac, #16573, RAM thread, #10771, slow-motion thread.