DARASK Lora Loader

Stacking five LoRAs in stock ComfyUI means five LoraLoader nodes and a wiring fan. The DARASK Lora Loader collapses that into one node with one row per LoRA - toggle, name picker, strength - which is the exact interaction model of rgthree's Power Lora Loader, the pack the community treats as the default for this job. It's an independent implementation with a trick up its sleeve: it reads rgthree-saved workflows, so you can load an old rgthree graph and the LoRA stack just works.

How it works

+ Add LoRA appends a row. Each row has a drag handle to reorder, a toggle to enable/disable, a LoRA picker that shows your models/loras/ tree as submenus, and a model strength that you scrub or type. Right-click a row for move up/down, enable/disable, and a Show CLIP strength option that reveals strengthTwo - the separate strength applied to the CLIP side, which matters for LoRAs trained with both model and text-encoder components. Rows with on = false, "None", or both strengths at zero are skipped; rows without an explicit strengthTwo apply strength to CLIP too, matching rgthree behavior.

Model and CLIP sockets are both optional - connect only model and it still works, which is how it slots between a UNETLoader and the sampler in the Anima stack.

Two implementation details are worth knowing because they explain why it "just works" with your existing setup. First, the widget storage format is the same dict-per-LoRA format rgthree uses, so a workflow saved with Power Lora Loader loads straight in. Second, LoRA file names resolve by fuzzy match - exact, then basename without extension, then substring - so metadata that names a LoRA loosely still finds the file.

What it's for

Beyond hand-stacking, this is the LoRA side of the pack's EXIF story: the DARASK Exif Apply nodes recognize this loader's format when they scan embedded workflows, so LoRAs recovered from an image's metadata can be re-applied cleanly. In the Anima example workflow, the recommended order is UNETLoader → DARASK Lora Loader → Step Cache → Sampling Tuner → KSampler.

Install

Part of the DARASK pack: ComfyUI Manager → search DARASK → install, restart. No models or extra dependencies.

Gotchas

Two things bite people. First, if the model input isn't connected, this node silently passes None through - and downstream nodes like the Step Cache or Sampling Tuner then throw 'NoneType' object has no attribute 'clone'. The loader is permissive; your wiring has to be right. Second, don't expect the rgthree feature set wholesale - this is the row-stacking interaction, not the trigger-word lookups or CivitAI integration that Power Lora Loader is also famous for. If you live in those, keep rgthree; if you want the pack's EXIF story to include your LoRAs, use this one.