Bubba LoRA Loader

Bubba LoraLoader is the standard LoRA loader - pick a LoRA, set a model strength and a clip strength, apply it to the model and CLIP - with one addition that earns it a place in the pack: it appends the LoRA's name to the pipe's metadata. So when Bubba Save Image writes bubba_metadata into the PNG, every LoRA that touched the image is listed. The core loader forgets; this one keeps the receipt.

If that sounds familiar, it's the same philosophy as every other node in this pack: provenance that travels with the image. You've applied LoRA X to a hundred images, and months later you're looking at a favorite and want to know what made it - the metadata answers, including the LoRA list. It's the difference between "I think it was the hair-color LoRA" and knowing for sure.

The inputs that matter

• lora_name - dropdown of every LoRA in your models/loras folder.
• strength_model - how strongly the LoRA affects the denoising model. Default 1.0, range -100 to 100.
• strength_clip - how strongly it affects the text encoder. Also default 1.0.

The two strengths are the part people get wrong. A LoRA isn't one knob; model strength changes the image generation, clip strength changes how the prompt is interpreted. On anime character LoRAs you often want model strength near 1.0 and clip strength lower (0.6–0.8), because the clip side is what makes the prompt over-obey the character's tags. Negative strengths are legal and do something - they push against the LoRA - but they're a last resort, not a tuning range. If you're starting cold: keep both at 1.0, then adjust model strength to taste.

The pipe contract: optional pipe in, and it applies the LoRA to the pipe's model/CLIP if you don't wire them explicitly. Explicit model and clip inputs override the pipe. Outputs are the updated pipe, updated metadata, the patched model and clip, and a lora_name string.

One LoRA at a time, by design

This node applies exactly one LoRA. If you stack multiple Bubba LoraLoaders in a chain, each one patches the model and CLIP from the previous and appends its name to the metadata - that works, but the chain gets long, and the pack has a better answer for that: Bubba LoRA Stack, which applies up to six LoRAs in slot order from a single node, with per-slot enable toggles, and records all of them. The rule of thumb: one LoRA, or a couple you want to think about individually → this node; a standing character+style+quality combo → the Stack.

Install

Standard pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/bubbafett5611/Bubba_Nodes

or ComfyUI Manager → search "Bubba Nodes", restart. ComfyUI 0.27.0+, no extra dependencies, no downloads beyond the LoRA files themselves.

The classic gotcha applies here as always: the dropdown only lists what's in models/loras, so a fresh download needs to land in the right folder and (in some setups) a refresh before it appears. And remember the clip/strength asymmetry - if a LoRA "isn't doing anything," the usual cause is a near-zero strength on one side, not a broken node.