LoRA Online

LoRA Online is the node that says "paste the link, I'll fetch it." Give it a direct URL to a LoRA file - or a Civitai model page URL - and it downloads the file, applies it to your model with a strength you set, and hands back the merged model. No manual download, no dragging files into models/loras, no refresh. For anyone who's ever received a workflow that references a LoRA they don't have, this is the node that makes that workflow runnable in one step.

How it works

The download is asynchronous (via aiohttp), which means ComfyUI stays responsive while a big safetensors file streams in - and the node reports progress through ComfyUI's execution progress API, so you can see it working. A few mechanism details worth knowing:

• Civitai URLs are resolved automatically. Paste https://civitai.com/models/123456 (optionally with a ?modelVersionId= to pin a specific version) and the node hits the Civitai API to find the actual download link before fetching. Direct .safetensors links work as-is.
• It saves into a dedicated subfolder - ComfyUI/models/loras/Flow-Assistor-LoRA/ - so it won't scatter files into your curated LoRA collection. The pack even registers an HTTP route with an "open folder" button in the UI.
• save_model defaults to on (keep the file). Flip it off and the LoRA is deleted after the run - useful for one-off tests you don't want to keep. force_redownload re-fetches even if a cached copy exists.

Inputs: model (your checkpoint model), url, strength_model (−10 to 10, default 1.0), save_model, force_redownload. Output: model - the model with the LoRA merged in, wired to your sampler exactly like a normal LoRA loader's output.

The honest take

This is a genuinely nice idea with a caveat attached: it only works when the source is reachable and permissive. Civitai links generally resolve fine, but some creators gate downloads behind accounts or rate limits, and a flaky host means the download fails - the node handles failure gracefully (it logs and passes the model through unmodified rather than crashing), which is the right behavior but means you might silently run without the LoRA. And remember the LoRA merge rules from the training docs don't change just because the file came over the wire: architecture lock-in still applies, and a wrong-architecture LoRA will misbehave exactly the same way it would loaded locally.

Installing

Part of ComfyUI-Flow-Assistor - ComfyUI Manager (search "Flow Assistor") or:

cd ComfyUI/custom_nodes
git clone https://github.com/Merserk/ComfyUI-Flow-Assistor.git

Restart after cloning. Current ComfyUI required (V3-only pack). No extra pip dependencies - aiohttp ships with ComfyUI - but obviously you need working network access from the machine running ComfyUI.

Where people get burned

The silent-failure path is the big one: if the download or the merge fails, you get the original model back and (unless you're watching the console) no obvious sign anything went wrong. Watch for the [LoRA Online] console lines. Also, cached-file behavior: with save_model on, the same URL won't re-download unless force_redownload is set, so a re-run after the file was updated remotely will use the stale local copy. And keep strength_model sane - this is a real merge, and a strength of 2 on a LoRA trained for 1 will give you the same artifacts it would locally.