Cloud Lora Loader

If you've used ComfyUI locally, you already know the LoraLoader node: it takes a model, a CLIP, and a LoRA file, and returns a patched pair. Cloud Lora Loader is that node, translated for the cloud. You wire it between a Cloud Checkpoint Loader and your sampler, pick a LoRA from a dropdown, and the LoRA - which lives on Comfy Cloud, not on your disk - gets applied inside the cloud job. You never download the weights, and the image comes back through the terminal node like nothing out of the ordinary happened.

The pack is a proof of concept that assembles a workflow JSON and ships it to ComfyUI Cloud; this node contributes a LoraLoader entry to that JSON. It takes the CLOUD_MODEL and CLOUD_CLIP handles you've built up so far, and returns patched handles for the rest of the chain. Locally a LoRA of a few GB means disk and RAM; cloud-side it means one dropdown pick.

The inputs that matter

• model and clip - the CLOUD_MODEL and CLOUD_CLIP handles from your cloud checkpoint or UNET loader. Nothing to configure here.
• lora_name - a dropdown of the LoRAs the cloud actually has. This is the one that bites beginners: until you've populated the cache, it shows a single placeholder, <run Cloud List Models (loras) to populate>. See the gotchas.
• strength_model and strength_clip - both default to 1, range −10 to 10. Same as local: strength_model is how hard the LoRA rewrites the diffusion weights, strength_clip how hard it touches the text encoder. A "model only" LoRA (like many video ones) wants strength_clip at 0.

Outputs are a fresh model and clip pair of cloud handles - feed them into your Cloud KSampler, CFG guider, or whatever sampling path you've built.

Installation

The node ships in the ComfyUI-CloudAPI-worker pack:

cd ComfyUI/custom_nodes
git clone https://github.com/Dobidop/ComfyUI-CloudAPI-worker

Copy config.json.example to config.json and paste an API key from https://platform.comfy.org/profile/api-keys, then restart. It's in ComfyUI Manager as "ComfyUI-CloudAPI-worker" too. Dependencies are just requests, Pillow, and safetensors - the LoRAs live cloud-side, so there's no download step.

Where people get burned

• The dropdown is empty on first install. The LoRA list is cached in config.json from a background prefetch at startup. If it's empty, run a Cloud List Models node with folder=loras, then restart ComfyUI. If you try to queue with the placeholder still selected, the node raises No LoRA selected. Run Cloud List Models (folder=loras) and restart ComfyUI. - that's not a crash, that's the pack telling you to do the housekeeping step.
• Cache is 24h stale. A LoRA you upload to Comfy Cloud won't show up until the next prefetch cycle. Same fix: Cloud List Models + restart.
• Don't confuse this with CloudLoraLoaderModelOnly. If your model has no text encoder attached (common with video and distilled checkpoints), there's no clip socket to feed - that variant exists exactly for that case.