◎ Radiance LoRA Stack

A serious workflow with three LoRAs is already a wiring mess - three LoraLoader nodes, six wires, and every strength tweak means finding the right slider on the right node. Radiance LoRA Stack collapses that into one node with five LoRA slots, each with its own model and CLIP strength, and outputs a single LORA_STACK object. The pack's docs frame it as "multi-layer LoRA management in a single node," and that's exactly the plumbing win: one node, five stacks if you chain them, and a clean socket you feed into the pack's unified loader.

If you're coming from the stock LoraLoader, the mental shift is that you're not applying LoRAs to a model directly - you're composing a stack that the loader consumes. That indirection is what makes the chaining work.

How it works

Each of the five slots - lora_1 through lora_5 - has a dropdown listing the LoRAs in your models/loras folder (plus "None" to skip the slot), a lora_X_model strength, and a lora_X_clip strength. The stack accumulates them in order. Two details make it more than a five-input bundle:

• Chaining. There's an optional lora_stack input on top. Feed an upstream LoRA Stack into it and the stacks concatenate - the docs say "chain multiple stacks together," so five slots is a floor, not a ceiling. One node per group, one wire per stack.
• Per-LoRA model vs CLIP. Model strength controls how hard the LoRA applies to the diffusion model; CLIP strength controls how hard it applies to the text encoder. For most LoRAs you keep them equal, but the split is there for the ones that only want to touch one side.

The output is a LORA_STACK - not a model. It feeds the Radiance Unified Loader, the pack's smart model loader (Flux/SDXL with precision auto-detection), which applies the accumulated stack. That loader is the intended consumer; the stack type is the pack's own convention, so you won't be plugging it into a stock Load Lora node.

The inputs that matter

• lora_stack - optional upstream stack to chain.
• lora_1..lora_5 + their model/clip strengths - the meat. Defaults: LoRA "None", model 1.0, clip 1.0. Strengths range -2 to 2, so you can also run a LoRA negative - a legit trick for style removal.

Output: one lora_stack.

Why the pack's loader matters here

The KB's LoRA training doc is where you'll find the "how to make a good LoRA" half; this node is the "how to actually use a pile of them" half. The Radiance Unified Loader reads this stack and handles Flux vs SDXL automatically - which is the real convenience, because stacking five LoRAs manually with the right precision per model is exactly the kind of fiddly work that eats an afternoon. That said, this node is a Radiance convention: if you're not using the Radiance loader, the stack output is useless to you, and stock LoraLoader chaining does the same job with more nodes and more visual clutter.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/fxtdstudios/radiance.git
cd radiance
pip install -r requirements.txt   # or requirements_windows.txt / requirements_linux.txt / requirements_mac_silicon.txt

Or search Radiance in ComfyUI Manager, restart. Linux needs libopenexr-dev first.

Gotchas

The stack isn't a model - it's a spec. Wire it into anything but the Radiance loader and you'll get a type mismatch. Second, stacking LoRAs is additive, not magical: five style LoRAs at 1.0 each will fight each other and probably produce a mess. The community wisdom is to treat stack slots as a reason to lower individual strengths, not to stack five full-strength LoRAs. Third, the dropdowns list whatever's in models/loras at node-build time - drop a new LoRA in the folder and you may need to reload the node or restart to see it.