Wan Model Loader

Here's the nicest thing about the flow2-wan-video pack: Wan Model Loader (WanVideoModelLoader_F2) is the only model-loading node you need. No separate checkpoint loader, no text-encoder loader, no VAE loader, no CLIP-Vision loader. One node drags in the entire Wan stack - the diffusion model and its companions - and hands you a ready MODEL for the patcher.

How it works

The unet_name dropdown is the interesting part. The list isn't hardcoded to local files - the pack queries HuggingFace at startup and builds the list live from three sources: city96's GGUF repos (Wan 2.1 I2V 480p/720p, T2V 14B, and the Fun-InP inpaint model), Comfy-Org/Wan_2.1_ComfyUI_repackaged (fp8 and BF16 diffusion files), and Kijai/WanVideo_comfy (the e5m2 fp8 quants). That's why you'll see dozens of wan2.1-fun-14b-inp-Q4_K_M.gguf-style entries and a pile of BF16/fp8 names - about 89 choices at last count.

Pick a name and the node:

• downloads it into models/diffusion_models on first use (this is the big download - a 14B GGUF is 8–12GB, so budget the disk and the patience),
• loads GGUF files with a bundled, city96-derived GGUF loader (GGML ops, on-the-fly dequantization) and plain diffusion files with ComfyUI's own loader,
• and then auto-downloads the four companions into their folders: the umt5_xxl fp8 text encoder, clip_vision_h, the wan 2.1 VAE, and the taew2_1 TAE used for live previews. You never fetch those by hand, and you never wire a separate CLIP or VAE node.

It also handles up to three LoRAs. lora_1/2/3 each have a _strength (range −10 to 10, 1.0 default) and the node caches loaded LoRA state dicts per slot, so swapping strengths mid-session doesn't re-read files every run. The 2.1-trained character LoRAs carry forward to 2.2, which is one of Wan's best features - you'll actually use this.

Which model should you pick?

Wan 14B is the quality pick, and GGUF is how most people run it. The community ladder applies directly: Q8 is basically fp16 at half the size - if it fits, use it; Q5_K_M is the accepted sweet spot for 12GB cards; Q4 is the "noticeable but playable" compromise. 24GB VRAM is comfortable, 12–16GB is the floor with Q5, and if that's tight, the 1.3B T2V models exist so you can still iterate. Want the Fun-InP inpaint workflow (start frame + end frame)? That's the wan2.1-fun-14b-inp-* family.

The output is a single MODEL, which goes into the Wan Model Patcher's model input.

Installing

This is the whole pack again: ComfyUI Manager → search flow2-wan-video, or

cd ComfyUI/custom_nodes
git clone https://github.com/Flow-two/flow2-wan-video.git
pip install -r requirements.txt

Restart, then let the loader pull its files. Requirements are genuinely light (gguf, sentencepiece, protobuf); the heavy lifting is the downloads, not the install.

Troubleshooting

First run will sit there "downloading" for a while - that's normal, it's pulling the model and the encoder. If the dropdown is empty, the live HuggingFace query failed and the pack fell back to a local unet_gguf folder list; put a GGUF file in models/unet or models/diffusion_models and it shows up.

The pack-wide warning applies here more than anywhere: flow2-wan-video patches global ComfyUI internals when it loads, and community threads repeatedly trace broken native Wan workflows back to this pack being installed (the telltale Given groups=1, weight of size [5120, 36, 1, 2, 2] channel error). If your other Wan workflows break after installing this, move the pack out of custom_nodes - the models it downloaded stay put and native ComfyUI will use them fine.