ADMultiLoraCombiner

ADMultiLoraCombiner is the answer to a question you'll hit the moment you use this pack's SD sampler: it takes one AD_LORA input, and you have three LoRAs you want on at once. The combiner merges up to four AD_LORA wires into one, and that one goes into the sampler. It's the AsyncDiff pack's version of a LoRA stack, and it's about as close to frictionless as anything in this pack gets.

The inputs are all optional, which tells you the design: lora_1, lora_2, lora_3, lora_4, each an AD_LORA. Hook up two, three, or all four - whatever you have. The output is a single AD_LORA that carries the combined effect, and it plugs into the lora input on ADSDSampler. That's the whole workflow: ADLoraLoader per file, ADMultiLoraCombiner to merge, sampler to generate.

Why this node exists instead of you just chaining loaders

Because the pack's sampler accepts exactly one LoRA input, and there's no good way to make two LoRA loaders share a wire. Combining before the sampler is the standard pattern - it's the same idea behind the Power Lora Loader in rgthree that the ComfyUI community leans on, just packaged for the pack's own AD_LORA type. Merging up to four into one wire also keeps your graph readable; you get one clean input instead of a pile of adapter nodes crowding the sampler.

A couple of practical notes:

• Order matters. LoRA applications are not commutative - the order you stack them changes the result. The combiner applies lora_1 through lora_4 in order, so put your most foundational LoRA first (the style or character base) and the finishing touches later. If the output looks mushy, swap the order before you start tuning weights.
• Weights still live on the loaders. Each ADLoraLoader has its own weight slider, and the combiner just merges whatever it receives. If a LoRA is fighting the others, drop its weight at the loader, not after.
• Four is the ceiling. More than four LoRAs and you need to combine in stages or reconsider what you're trying to do - at some point you're not stacking LoRAs, you're making soup.

Install

Standard for the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/SlackinJack/asyncdiff_comfyui.git
cd asyncdiff_comfyui
./clone_asyncdiff_repo.sh   # pulls the upstream AsyncDiff framework - required

Restart ComfyUI (or use ComfyUI Manager, search "asyncdiff_comfyui").

Where people get burned

The usual suspects apply: this is a multi-GPU pack, so without multiple cards the sampler underneath is slower than stock ComfyUI rather than faster. And the wiring is narrower than it looks - the combiner's output only feeds the SD text-to-image sampler's lora input; the video and upscale samplers don't take LoRAs at all. Keep your expectations set by the pack's support note too: the author's stated working path is stable-video-diffusion-img2vid, and the SD lane this combiner serves is the less-tested side of the house. It works - you're just on the frontier.