Cloud UNET Loader

The UNET Loader in vanilla ComfyUI is the node you use when you don't want a checkpoint with a CLIP and VAE bolted on - just the diffusion model file, thank you. Cloud UNET Loader is that same node, and the "diffusion model file" lives on ComfyUI Cloud's storage, not your disk. For Wan, Hunyuan, LTX Video and Flux workflows this is usually the right way to load the model, because those models ship as standalone diffusion files rather than checkpoint bundles.

How it works

Same pack trick as every other Cloud node: nothing downloads and nothing runs locally. The node takes your pick from the cloud's diffusion_models folder and appends a UNETLoader node to the assembled workflow JSON. The actual load happens on the cloud instance when a terminal node finally submits the job. What you get back is a CLOUD_MODEL handle that threads through your graph - conditioning, guider, sampler - until the chain is complete.

The inputs and outputs that matter

• unet_name - a dropdown populated from the cloud's diffusion_models folder. It'll show a placeholder (<run Cloud List Models (diffusion_models) to populate>) until the cache is filled.
• weight_dtype - default, fp8_e4m3fn, fp8_e4m3fn_fast, or fp8_e5m2. This picks the precision the cloud instance loads the model at. fp8 is the classic "fits and loads faster" choice for the big video models - a 14B Wan is a chunk of VRAM even for the cloud - while default keeps whatever the model file specifies.
• Output: model (CLOUD_MODEL) - wire it into a model-sampling or guider setup and on to the sampler.

Why you'd reach for it

Two reasons. First, video models: the pack's own example workflow (video_wan2_2_14B_i2v_cloud.json) loads Wan 2.2 14B as a diffusion model this way, with a separate Cloud VAE Loader and text encoders alongside. Second, the entire point of this pack: you get to run a 14B model you couldn't fit on your own card, because "your GPU" is an A100 sitting on ComfyUI Cloud. Wan 2.2 is the last open-weight Wan, so this is about as good as that trade gets.

Installing it

Install via ComfyUI Manager (search "ComfyUI-CloudAPI-worker") or:

cd ComfyUI/custom_nodes
git clone https://github.com/Dobidop/ComfyUI-CloudAPI-worker
cd ComfyUI-CloudAPI-worker
cp config.json.example config.json

Paste your API key from https://platform.comfy.org/profile/api-keys into config.json and restart ComfyUI. Only requests, Pillow and safetensors are needed locally - the model file itself never touches your machine.

Common issues & troubleshooting

• Empty dropdown is the one everyone hits - run a Cloud List Models node with folder=diffusion_models, then restart ComfyUI so the dropdown picks up the fresh cache. The startup prefetch refreshes on a 24-hour cadence in the background, so it fixes itself overnight but not instantly.
• The node raises a "No diffusion model selected" error - that's the guard firing when unet_name starts with <. The message tells you exactly what to do (run Cloud List Models, restart). Read it before blaming the node.
• Proof of concept - if a cloud model folder you expect isn't listed, it may genuinely not be on the cloud instance's list yet. The README's status section is upfront that the node coverage is a work in progress.