ComfyUI Wan 2.2: how to run it locally, VRAM needs and GGUF

Here is the setup that gets Wan 2.2 running in ComfyUI: model choice, folders, VRAM figures and file sizes for every format. Everything comes from the maker's repo, ComfyUI's docs and the Hugging Face file listings, checked on 2026-10-26.

Which Wan 2.2 model should you run in ComfyUI?

Wan 2.2 has a 5B model for low VRAM and two 14B models (text-to-video and image-to-video) for higher quality.

The 14B models are a mixture of two experts, 27B parameters in total with 14B active per step. The high-noise expert lays out the scene and the low-noise expert refines detail. That is why a 14B workflow has two model loaders (Wan team, GitHub README).

All three are Apache 2.0. To swap characters in existing footage instead, see the Wan 2.2 Animate guide. For hosted newer versions that have no downloadable weights, see the Wan 2.5 guide.

How do you set up Wan 2.2 in ComfyUI?

Update ComfyUI, open a built-in Wan 2.2 template, and download the model files it names into the right folders.

1. Update ComfyUI. ComfyUI's docs say to use the Nightly version if nodes are missing.
2. Open Templates and choose a Wan 2.2 workflow: 5B Video Generation, 14B Text to Video, 14B Image to Video or 14B First-Last Frame.
3. Download the files from the Comfy-Org/Wan_2.2_ComfyUI_Repackaged page on Hugging Face.
4. Place them as below, then restart ComfyUI and run the template.

The docs list wan2.2_ti2v_5B_fp16 for 5B, fp8 scaled files for 14B text-to-video and fp16 files for 14B image-to-video. The 14B image-to-video steps in the docs name the text-to-video fp8 files, so make sure the file names in your loader nodes match the task you run.

How much VRAM does Wan 2.2 need?

Official figures disagree because they describe different setups: 8 GB for 5B in ComfyUI, 24 GB for 5B at 720p and 80 GB for 14B in the Wan repo.

ComfyUI offloads weights to system RAM, so the 14B models run on cards far below 80 GB, at the cost of speed. The Wan README also carries a GPU table of generation time and peak memory under its stated settings, plus one headline figure: a 5-second 720p clip from 5B in under 9 minutes on one consumer GPU. Plan for system RAM too: the two 14B files plus a text encoder total tens of gigabytes.

How big are the Wan 2.2 14B files in fp16, fp8 and GGUF?

The two 14B experts total 57.2 GB in fp16, 28.6 GB in fp8 and between 10.6 and 30.8 GB in GGUF, before the 6.74 GB fp8 text encoder.

This table adds the high-noise and low-noise file for image-to-video (T2V sizes match for fp16 and fp8). Sizes are from the Hugging Face listings on 2026-10-26.

Other files: text encoder umt5_xxl_fp8_e4m3fn_scaled 6.74 GB (fp16 version 11.37 GB), wan_2.1_vae 0.25 GB, wan2.2_vae 1.41 GB, 5B diffusion model 10.0 GB.

Disk size is not VRAM use. The sampler runs the high-noise expert first and the low-noise expert second, so how much sits in VRAM at once depends on your offloading settings. Use the table to rank formats.

Worked example: a card with 16 GB of VRAM cannot hold both fp8 files (28.59 GB), but Q4_K_M files total 19.30 GB and each single expert is 9.65 GB, which is the case that makes GGUF attractive on 12 to 16 GB cards.

How do you use Wan 2.2 GGUF in ComfyUI?

Install the ComfyUI-GGUF node, put the GGUF files in models/unet, and replace the Load Diffusion Model node with the GGUF Unet loader.

• Clone city96/ComfyUI-GGUF into ComfyUI/custom_nodes and run pip install --upgrade gguf (the node's README). The node is Apache 2.0.
• Download one high-noise and one low-noise file from a repo such as QuantStack/Wan2.2-I2V-A14B-GGUF. Files are direct conversions of the original model and keep its license terms.
• Open a 14B template, swap both model loaders for the GGUF loader and select the matching files.
• LoRA loading in this node is marked experimental, so test your speed LoRA on one clip before a long queue.

The QuantStack card lists Q2_K to Q8_0, and the ComfyUI-GGUF README notes that transformer models tolerate low bit rates better than convolution models. To pick your quant, run the same prompt and seed at Q8_0 and at the quant you can fit, then compare faces and fast motion.

How do you make Wan 2.2 faster in ComfyUI?

The lightx2v Lightning LoRA cuts sampling to 4 steps with no CFG, and its card claims a 20x speed-up.

ComfyUI's docs link a 4-step LoRA for 14B text-to-video. The Lightning card ships ComfyUI workflows for image-to-video and text-to-video, and says version 1.1 of the text-to-video LoRA is slightly better than version 1. The Comfy-Org repackaged repo holds the LoRA files at 1.23 GB each (a high-noise and a low-noise file per task).

The 20x figure is the maker's claim. The card flags artifacts in scenes with extremely large motion and occasional reversed vehicle direction.

Is running Wan 2.2 locally cheaper than a hosted API?

Local runs cost electricity and your time, while hosted models charge per second, so the answer depends on how many clips you keep.

Local generation has no model usage fee, but each render still costs electricity and time. Hosted APIs skip local installation, offer newer models and audio, and charge for generation under each provider's billing rules. Compare with the per-second rates in the AI video cost calculator and the Kling pricing guide. Wan 2.2 is a good fit for drafts and tests, and the AI video model finder can help you pick a hosted model for finals. For a final clip of a real person, get consent first.