Auto Download ALL WAN Models from Kijai's repo

If you run Wan through Kijai's ComfyUI-WanVideoWrapper - and a huge slice of the community does, because it was the first way to get SageAttention and TeaCache working properly on Wan - then you're downloading weights from Kijai/WanVideo_comfy. It's the repo where Kijai's own workflows point, including his fp8 repackaged files. This node is an eighteen-checkbox downloader for exactly that repo, so your wrapper workflow's loaders stop reddening with "model not found."

The idea is the same as the sibling Comfy-Org node: tick the files you want, run the workflow, and the downloader drops each one into the right models folder. The difference is the source - Kijai's naming and his fp8 variants (he ships both e4m3fn and e5m2 flavors) differ from Comfy-Org's repackaged layout, and you need the Kijai copies specifically if you're following his wrapper workflows or his Fun/Control builds.

What the checkboxes map to

Each boolean maps to a fixed filename in Kijai/WanVideo_comfy, downloaded to hardcoded folders under /workspace/ComfyUI/models/:

• diffusion models (the Wan2_1-T2V-14B-*, Wan2_1-I2V-14B-480P/720P-*, the 1.3B T2V, plus Wan2.1-Fun-Control-14B and Wan2.1-Fun-InP-14B) → unet/
• VAEs (Wan2_1_VAE_bf16 / fp32) → vae/
• text/vision encoders (umt5_xxl_enc_*, open-clip-xlm-roberta-large-vit-huge-14_visual_*) → clip/
• taew2_1 (Kijai's WAN Animate weight) → unet/

For a baseline Wan 2.1 T2V on a 12 GB card you want the 1.3B T2V model (or the 14B fp8_e4m3fn), the umt5_xxl_enc_fp8_e4m3fn, and one VAE. Note there's no fp8-scaled option here - that's the Comfy-Org split that carries _fp8_scaled; Kijai's repo splits its fp8 into e4m3fn (good quality/size balance) and e5m2 (less accurate mantissa, slightly smaller). If a workflow says e5m2, tick e5m2; don't let the loader silently substitute.

Install and the shared trap

Same pack-wide install - ComfyUI Manager search ComfyUI_AutoDownloadModels, or:

cd ComfyUI/custom_nodes
git clone https://github.com/AIExplorer25/ComfyUI_AutoDownloadModels
cd ComfyUI_AutoDownloadModels
pip install -r requirements.txt

then restart.

The gotcha is identical to the Comfy-Org node and worth repeating loudly: everything hardcodes /workspace/ComfyUI/models/. The author's frame is a RunPod-style cloud box where /workspace is home. On a local machine the downloads land somewhere ComfyUI never looks, or fail outright. Make /workspace/ComfyUI/models point at your real models directory:

ln -s /path/to/ComfyUI/models /workspace/ComfyUI/models

The node's output is a path string, but it's really a status echo - the code returns only the last file's path after a loop, not all of them, so don't chain it into anything that tracks every model. The queue blocks while a big file downloads; that's the node working, not a crash.

Who this is for

People running Kijai wrapper workflows or his Fun (control/inpainting) builds, on a cloud instance or with the symlink trick. If you're on native ComfyUI's own Wan 2.1 examples, the Comfy-Org repackaged sibling node is the more natural match - different repo, same checkbox UX. Both are honest about being download conveniences: they don't install the wrapper, don't build a workflow, don't check VRAM. Just get the weights where the loaders want them.