⭐ Star 3 LoRAs

Anyone who has stacked LoRAs in ComfyUI has met the failure mode: LoRA A applied at strength 1, then LoRA B applied on top, and suddenly the whole image has the same face, or the style saturates into mush. The reason is that LoRA order and strength both matter, and doing it by hand means a chain of LoraLoader nodes where a wrong connection silently changes everything. Star 3 LoRAs (Star3LoRAs) is the pack's answer: up to three LoRAs in one node, each with an independent model strength and CLIP strength, applied in a fixed, readable order, plus a lora_stack output that feeds the pack's starters for internal conditioning.

How it works

It's a sequential loader: lora1_name applies first, then lora2_name, then lora3_name - order is fixed and visible, which is exactly what you want because the last LoRA in the chain has the most influence over the final weights. Each has two strength sliders (range -100 to 100, default 1.0):

• strengthN_model - how hard it pushes the diffusion model. This is the number you actually tune for style/concept.
• strengthN_clip - how hard it pushes the CLIP text encoder. Set to 0 to apply a LoRA only to the model and leave prompt interpretation untouched - the trick for mixing a style LoRA without it dragging the prompt semantics around.

The tooltips call it out plainly: pick "None" to skip a slot, use the strengths to modify model/CLIP respectively. Negative strengths invert the LoRA's effect, which is occasionally exactly what you want (removing a baked-in style).

The inputs and outputs

Everything is optional - you can even drop a model in with no LoRAs and just get the passthrough back. The real shape:

• model (and optional clip) - what the LoRAs get applied to.
• LoRA_Stack input - an existing stack, so you can chain another LoRA node in front of this one.
• lora_stack output - the star of the show. Wire this into FluxStartSettings / SDXLStartSettings / SD35StartSettings's LoRA_Stack input and the starter applies all three LoRAs internally, to model and conditioning, in one shot. This is why the node's description says "Can be connected to FluxStart for internal conditioning."

Outputs: model, clip, lora_stack.

Where it sits in the pack

This is the pack's answer to the community-standard "power LoRA loader" idea - the rgthree Power Lora Loader is the famous one in that space, and Star 3 LoRAs is the same concept dialed down to exactly three slots with per-LoRA model/CLIP separation. If you only ever stack two LoRAs, this is overkill but harmless; if you regularly stack three, it beats a chain of four nodes. The per-slot CLIP strength is the feature the flat stackers lack.

Real-world notes

The usual LoRA wisdom still applies on top of the plumbing: strengths around 0.6–0.8 for style LoRAs, 0.8–1.0 for character/concept, and if two LoRAs fight each other, lowering both a notch beats zeroing one. Because everything's optional, the dropdowns show your full models/loras folder - a LoRA listed but set to "None" is skipped, so you can wire a template with three slots and only use two.

Install

Part of the StarNodes pack:

cd ComfyUI/custom_nodes
git clone https://github.com/Starnodes2024/ComfyUI_StarNodes

or ComfyUI Manager → search Starnodes, restart, find it under ⭐StarNodes/Sampler. No extra dependencies. Note the strength ranges go to ±100 - the sliders are permissive on purpose, but treat anything beyond ±2 as "experimenting with broken output."