Visual LoRA Picker+

You know the feeling: four hundred LoRAs in a dropdown, and half the filenames are xl_more_detailed_face_v1 and style_anime_final_v3_FINAL. Visual LoRA Picker+ replaces the scrolling with a thumbnail browser - pick the image that matches the LoRA you actually want, and the node resolves the file for you. If you keep preview images next to your LoRAs, this is the most pleasant way to select one in ComfyUI.

It's a small node from the same pack as VACE Stitcher - ComfyUI-FBnodes, a grab-bag of quality-of-life tools by FranckyB - and it plays in the same space as rgthree's Power Lora Loader. The difference: Power Lora Loader is a full stack manager with trigger-word lookup; this is a focused "pick one by sight" node that slots into any LoRA workflow, for SDXL, Flux, Wan, whatever.

How it works

The mechanism is almost embarrassingly simple, which is why it works. You open a file browser pointed at your LoRA folder, and it shows matching images (webp, jpg, png) that share the same path and basename as a .safetensors file. Select a thumbnail, and the node derives the LoRA by swapping the extension - my_style.png becomes my_style.safetensors - then resolves it to a ComfyUI-relative path inside your loras tree.

The result is a LORA_STACK: a list of (path, model_strength, clip_strength) tuples, the same stack format rgthree and other loaders use. If you feed in an incoming lora_stack, your pick gets appended to it; if not, you get a single-entry stack. That makes it trivial to chain: pick a base character LoRA, then a style LoRA, and hand the combined stack to an apply node.

One subtlety worth knowing: strength_model is a single slider, and the node sets both model and clip strength to that same value. That's fine for most cases - model strength is what you're usually tuning anyway - but if you want different model/clip strengths per LoRA, this isn't the node to do it with.

Inputs and output

• image - the selected thumbnail's path. Click the browser to choose one. There's also a (none) selection that passes an incoming stack through unchanged, handy when you're conditionally picking LoRAs.
• strength_model (default 1.0, range -100 to 100) - the strength applied to the picked LoRA.
• lora_stack (optional) - an incoming stack to append to.
• Output: lora_stack - wire this into an applier. If you're using this pack, that's Apply LoRA+ (same repo, takes a LORA_STACK into MODEL + CLIP). It'll also feed any other stack consumer.

Installing it

Install the whole pack once and you have every node in it:

cd ComfyUI/custom_nodes
git clone https://github.com/FranckyB/ComfyUI-FBnodes.git
pip install -r ComfyUI-FBnodes/requirements.txt

Or search "ComfyUI-FBnodes" in ComfyUI Manager. Dependencies are light - av for the pack's video nodes and color-matcher for the VACE stitcher - nothing special needed for this one. Restart ComfyUI after cloning.

The gotcha: the images are on you

Here's where people get burned. This node does not fetch previews from anywhere. CivitAI shows you a preview image on the model page, but that image doesn't live next to your downloaded LoRA unless you put it there. If you select a thumbnail with no matching .safetensors (or vice versa - a LoRA with no image beside it simply won't appear in the browser), you get a FileNotFoundError: Matching LoRA not found and the node fails.

The naming also has to match exactly, extension aside. my_style (1).webp beside my_style.safetensors won't resolve because the basenames differ. So the real workflow here is: organize once - drop a preview image next to each LoRA with the same name - and from then on selection is genuinely pleasant. For big libraries where you've been relying on filenames, that up-front cost is worth it; for a handful of LoRAs you can keep, it probably isn't.

That's the honest trade-off. It's a tool you either use constantly or never, and it's only as good as the folder you point it at.