Load LoRA (Online)

Load LoRA (Online) is the pack's single-LoRA node: pick an adapter from DoubTech's hosted catalog, and the node downloads it into your models/loras folder and applies it - no manual hunting for a file and hand-placing it. The interface mirrors ComfyUI's stock LoraLoader, so if you know that, you already know this.

The catch, and it's worth knowing before you build a workflow around it: the LoRA catalog is two entries. Right now the whole list is Detail Tweaker LoRA (the add_detail.safetensors file) and LowRA 2.0. Both are small utility adapters that live on Dropbox. So this node is less a serious workflow workhorse and more the "test whether the pack works" node - and there's a code-level wrinkle you should know about.

The inputs that matter

It takes the usual three plus strengths:

• model and clip - feed it the MODEL and CLIP outputs from your checkpoint loader, same as the stock LoraLoader.
• lora_name - the dropdown, populated from the catalog.
• strength_model and strength_clip - floats from -10 to 10, both defaulting to 1.0. The classic advice still applies: 1.0 is often too hot; 0.5–0.8 is the common sweet spot for both. With a Detail Tweaker-style LoRA, keep it low and let it do subtle work.

Outputs are MODEL and CLIP, wired between the checkpoint loader and the sampler exactly like the core node.

The mechanism

Same shared pack machinery: on startup it fetches the LoRA list from DoubTech's API, and on first use of a given LoRA it downloads the file into models/loras and loads it with ComfyUI's own load_lora_for_models. Once the file exists locally, no network. It also keeps an in-memory cache of the last-loaded LoRA so repeat runs skip re-parsing the file.

One honest caveat from reading the source: that cache has a bug. On the very first run - before anything is cached - the code path references a local variable that was never assigned, and it throws UnboundLocalError: local variable 'loaded_lora' referenced before assignment. I verified the logic in isolation, so if your first queue of this node dies with that exact error, that's what you're looking at. The workaround is trivial: use DTLorasLoader (the multi-LoRA node in the same pack, which doesn't have this bug), or use the stock LoraLoader against the file this node already downloaded for you. This is also a fair summary of this pack's QA level in general - it's MIT-licensed, by Yolan (DoubTech), last touched in late 2024, and it shows.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/yolanother/DTAIComfyLoaders

or search "DTAIComfyLoaders" ("Comfy UI Online Loaders") in ComfyUI Manager. Restart after. Only dependency is requests.

Where people get burned

• The default strengths. 1.0/1.0 out of the box is aggressive for most LoRAs. Drop to ~0.6 and see.
• First-run download stalls. No progress bar; the node sits there while it pulls the file from DoubTech's Dropbox link, which can also deliver an HTML page instead of the model. Small file in models/loras = the download went wrong, grab it manually.
• The catalog is tiny and frozen. Two LoRAs, SD-era, no refresh button - the list only updates on restart.

If you just want LoRAs that exist locally, the stock node does it better. This one's for poking at the online-loading idea without managing files.