Load Wan 2.2 Model

If you've read anything about running Wan locally, you've probably run into Kijai's ComfyUI-WanVideoWrapper - it's the de facto standard, the one most workflows on Civitai and Reddit are built around. LoadWan2_2Model is the other option: it's the loader for plain, un-retrained Wan 2.2 (T2V-A14B, I2V-A14B, TI2V-5B - the exact checkpoints Alibaba's Tongyi lab shipped) that ships inside VideoX-Fun, a separate node pack built by a different Alibaba team entirely. VideoX-Fun comes out of PAI, Alibaba Cloud's applied-AI group - the same team behind EasyAnimate - and their whole pitch is a self-contained integration: load and sample Wan without installing anything else alongside it.

Worth knowing going in: this pack's node menu category is still labeled CogVideoXFUNWrapper, a fossil from before it grew Wan support. Don't let that throw you when you're browsing the node list for a Wan node.

How it works

You pick a checkpoint name, VideoX-Fun assembles the diffusion transformer, VAE, and text encoder according to the matching config file, applies whatever GPU offload strategy you asked for, and hands back one bundled object - funmodels - that the sampler nodes downstream plug straight into. Nothing fancier than that; it's a standard load-and-configure step.

The inputs that matter

• model - which checkpoint to load: Wan2.2-T2V-A14B, Wan2.2-I2V-A14B, or Wan2.2-TI2V-5B (default T2V-A14B). This has to match a folder you've actually downloaded.
• config - a YAML file that has to correspond to the model family you picked (wan_civitai_t2v.yaml, _i2v.yaml, or _5b.yaml). Mismatch these and loading breaks.
• GPU_memory_mode - the speed/VRAM tradeoff. Per the README: model_cpu_offload moves the whole model to CPU after use (the recommended default if you have enough VRAM); model_cpu_offload_and_qfloat8 does the same plus quantizes the transformer to fp8, trading a little quality for more headroom; sequential_cpu_offload moves each layer individually - slowest, but the one that actually gets you through on tight VRAM. model_full_load is the fourth option and skips offloading entirely: fastest, but you need the VRAM to back it up.
• precision - fp16 or bf16.

Output is a single funmodels (FunModels type) that feeds into a Wan sampler node.

Installing it

Search ComfyUI Manager for VideoX-Fun first. It hasn't always shown up there reliably - community reports from early adopters describe having to install it by hand - so if Manager comes up empty, do it directly:

cd ComfyUI/custom_nodes
git clone https://github.com/aigc-apps/VideoX-Fun
pip install -r VideoX-Fun/requirements.txt

Restart ComfyUI after either path. Then get the weights: download the full HuggingFace repo for whichever checkpoint you're using (not just a .safetensors file - the whole folder, config files included) and drop it under ComfyUI/models/Fun_Models/, in a subfolder named to match the dropdown exactly, e.g. Fun_Models/Wan2.2-T2V-A14B/. The README says to budget 60GB+ free disk for this kind of work generally, and the A14B checkpoints are the heavy end of that.

Common issues

The model dropdown is empty or your download doesn't show up. Almost always a folder problem - wrong location (Fun_Models/ under models/, not the Diffusion_Transformer/ path the README also shows, which is for the standalone Python scripts, not ComfyUI), or a folder name that doesn't match the checkpoint string byte-for-byte.

Load fails with a config mismatch. Keep config paired to model: t2v checkpoint with the t2v yaml, i2v with i2v, 5b with 5b. Don't mix and match.

Out of memory on the A14B models. They're the 14B-active MoE tier - the same weight class the wider Wan 2.2 community runs on 24GB cards with care. Drop to sequential_cpu_offload, or step down to Wan2.2-TI2V-5B if you're on 12-16GB and just need it working.

You actually wanted Control or Camera features. This loader only gives you the stock checkpoints. If you want start/end-frame prediction, Canny/Depth/Pose control, or camera-motion control, you want LoadWan2_2FunModel instead - that's the PAI-retrained "Fun" variant family, and it's a different set of downloads.