🎨 Load Lora Filter

This is the pack's LoRA loader with two tricks the stock one doesn't have: it can download a LoRA on demand from a URL, and it ships with a curated dropdown of commonly-used LoRAs so you can grab one without hunting the internet for it. If you've ever loaded a workflow that wants a speed LoRA you don't have, this is the node that fetches it for you instead of erroring out.

The download behavior is the headline. Leave the Download_url field empty and it acts like a normal loader - you pick from the lora_name dropdown and it applies what's on disk. Fill Download_url with a Civitai or HuggingFace link and turn Download on, and it pulls the file into your loras folder (naming it from Lora_url_name, default model.safetensors) and uses it. That means a shared workflow can carry its own LoRA dependency as a link rather than an "install this first" note. The lora_name dropdown itself is a curated library of about 80 entries - a lot of them speed and utility LoRAs like Hyper-SD15-8steps, Flux_Turbo_8steps, the ACE+ set, Krea_lora - and if the file isn't already local, choosing it downloads it automatically. The fill field is a filter to narrow that long dropdown by keyword, which is where the "Filter" in the name comes from.

The inputs a beginner actually sets are few. Pick a lora_name (or feed a Download_url + Lora_url_name to grab one by link), then set strength_model and strength_clip - the two familiar strength dials, both defaulting to 1. The model and clip inputs are optional: pass them through and the LoRA is applied to them and handed back on the model and clip outputs, exactly like the standard LoraLoader. There's also a lora_path STRING output telling you where the file ended up, handy after an auto-download when you want to confirm what actually landed.

A practical note on the strengths, since this loader makes it easy to stack these: most of the curated entries are distilled step-reduction LoRAs (Hyper, Turbo, Lightning), and those want to run around strength 1 to hit their step counts - they're not "style" LoRAs you dial to taste. Treat them as switches more than knobs. Style and character LoRAs are the ones where strength_model in the 0.6–0.9 range is the usual play, and strength_clip you often leave at 1 or slightly under.

Install comes with the pack. ComfyUI Manager, search SDVN_Comfy_node; or cd ComfyUI/custom_nodes && git clone https://github.com/StableDiffusionVN/SDVN_Comfy_node, then pip install -r custom_nodes/SDVN_Comfy_node/requirements.txt from your ComfyUI root and restart. Because the download features lean on aria2c, Windows and macOS users need to install aria2c by hand for the auto-download to work; on Linux it's usually already present.

Troubleshooting is mostly download-side. If a Download_url fails, check that it's a direct model link the way Civitai/HuggingFace expose it and that aria2c is actually installed - a silent no-op on Windows is almost always the missing aria2c. If a curated pick doesn't download, same cause. And the usual LoRA-matching rule still applies: a LoRA trained for one architecture won't do anything useful on another - a Flux Turbo LoRA does nothing for an SDXL checkpoint, a Hyper-SD15 LoRA won't help Flux - so if you applied one and got zero visible change, that's the mismatch, not the node. Match the LoRA to your base model first, then worry about strength.