Auto Download ALL WAN Models from Comfy-Org/Wan_2.1_ComfyUI_repackaged

Wan 2.1 is where open-source video generation got serious - Alibaba's 14B diffusion transformer, Apache 2.0, an I2V variant the community loved. The catch has always been logistics: a working Wan 2.1 setup needs a diffusion model and a text encoder and a VAE, each in a specific folder, and ComfyUI's own example workflows point at Comfy-Org/Wan_2.1_ComfyUI_repackaged for all of them. This node is the pack's checkbox-driven installer for that exact repo: one boolean per file, tick what you want, run, and the weights land in the folders native ComfyUI loads them from.

The display name says "Auto Download ALL WAN Models" and that's half true - you still tick the files yourself, but you can tick all eighteen and get the entire repo's split-file layout in one go. That's the intended flow: stand up a fresh cloud box, tick everything, go make coffee.

How it works

Every checkbox is a separate hf_hub_download call against Comfy-Org/Wan_2.1_ComfyUI_repackaged. Unchecked, nothing happens for that file. Checked, the file downloads into a hardcoded destination:

• split_files/diffusion_models/* → /workspace/ComfyUI/models/unet/
• split_files/text_encoders/* → /workspace/ComfyUI/models/clip/
• split_files/vae/* → /workspace/ComfyUI/models/vae/
• split_files/clip_vision/* → /workspace/ComfyUI/models/clip_vision/

Because Comfy-Org split the repo to match ComfyUI's own folder conventions, the files are already in the right place for ComfyUI's Wan 2.1 loaders. You're skipping the "download to the wrong folder, then move it" dance entirely.

The checkboxes that actually matter

For a plain Wan 2.1 T2V run you need three things: one diffusion model, the umt5_xxl text encoder, and the VAE. Everything else is choice. The diffusion model toggles come in four flavors per variant - bf16, fp16, fp8_e4m3fn, and fp8_scaled - and this is where a little knowledge saves you 60 GB of disk:

• fp8_e4m3fn is the community default for 24 GB and 12 GB cards. It's the repackaged fp8 that ComfyUI loads natively without extra conversion.
• fp8_scaled is Kijai's scaled variant (from his own repo's fork of the format) - marginally better quality retention at the same size, but it's what his wrapper workflows reference.
• bf16/fp16 are the full-precision weights, ~30 GB each for the 14B models. You don't need these unless you're chasing every last bit of quality on a big card.

The 1.3B toggles are the small T2V model - the one to tick if you're on 8 GB VRAM and just want to test a workflow before committing to a 14B render.

Install and the one big trap

Install is pack-wide: ComfyUI Manager → search ComfyUI_AutoDownloadModels → install, or clone https://github.com/AIExplorer25/ComfyUI_AutoDownloadModels into custom_nodes/, pip install -r requirements.txt, restart.

Now the trap that will bite you the first time: the destination is hardcoded to /workspace/ComfyUI/models/.... That's the RunPod/cloud convention, which is what the author targets. On a normal local install, /workspace doesn't exist, and the node either errors out or quietly downloads gigabytes to a path ComfyUI never reads. Two fixes:

# Point the cloud path at your real models folder (Linux/macOS)
ln -s /path/to/ComfyUI/models /workspace/ComfyUI/models
mkdir -p /workspace/ComfyUI  # then the symlink above

Or run it on a cloud box and keep everything under /workspace as intended. Same story for the sibling Kijai and alibaba nodes - they share this hardcoding.

The output is a path string, but read it as a status signal: the code overwrites it each iteration, so you only get the last file's path back, not a list. Don't wire it into anything that needs to know where every model went. Downloads block the queue until done, which for the 14B fp8 files is many minutes - expected behavior, not a hang.

If all you need is one specific Wan file, the pack's plain AutoDownloadModels node is the cleaner tool. This one earns its keep when you're provisioning a new box and want the whole family in one run.