Lora Load from Subdirectory (Model Only)

This is the leaner sibling from the lorasubdirectory pack, and its display name says it all: "Lora Load from Subdirectory (Model Only)". Where the main LoraSubdirectory node takes MODEL and CLIP in and hands both back, this one takes just a MODEL and outputs just a MODEL. Same first-dropdown-filters-the-second idea, one less socket on every side. For its intended use that's not a loss, it's the point.

Why you'd reach for it

Two reasons, both worth having. First, a lot of modern LoRAs - Flux-lineage ones especially - patch the transformer, not the text encoder, so there's no CLIP side to modify. The extra socket on the full node is then just dead weight and one more wire to misroute. Second, a single MODEL output chains cleanly: stack several of these and each takes the previous MODEL, applies its LoRA, and passes the result on - no CLIP wire to keep in sync along the chain. If your workflow never touches CLIP after the checkpoint loads, this is the tidier tool.

The subdirectory dropdown works exactly like the parent node: it walks your loras folder tree, lists every directory that contains at least one .safetensors, and once you pick one, the LoRA list shows only that folder's files. Same filtering mechanism, same design note from the author that the directory can't be passed as a string input because the dropdown's array type drives the list.

The inputs that matter

• model (MODEL) - your checkpoint's output. Required.
• lora_directory - the new dropdown; pick the folder first.
• strength_model (default 1.0, −100 to 100) - how strongly to apply; negative values invert the effect.

One output: MODEL, wired straight into the KSampler (or into the next chained loader).

Install and the catch

Same pack, same install as the parent - ComfyUI Manager, search "lorasubdirectory", install, restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Hax0r778/lorasubdirectory

Zero dependencies, no model downloads, no install script; restart and it's loaded.

And now the part you actually need before you build around this. As shipped (version 1.0.1, a single initial commit), the Model Only variant's execute method calls a self.load_lora(...) helper that doesn't exist anywhere in the repository - so it raises AttributeError the moment you queue the graph, before it touches your model. The parent class it inherits from carries its own first-run bug too (loaded_lora never initialized in __init__). The dropdowns populate and look right; the execution code is unfinished. If you want this filtered-picking behavior working today, either patch nodes.py (init the cache and route the Model Only method to the parent's real loader), or use a loader the community actually runs - rgthree's Power Lora Loader stacks multiple LoRAs and is the one people reach for. The idea here is good; the code just isn't there yet.