LoRA Stack (Mango)

Loader (Mango) stops at five LoRAs, and if you hang around the community long enough you will eventually want more. Style LoRA for the look, a character LoRA for the face, a lighting LoRA, a quality LoRA - the stack piles up. LoRA Stack (Mango) is the node that sits between your base loader and your sampler and adds up to five more LoRA slots to whatever is already in the graph, including a stack that came from another Mango loader.

The neat part is that it doesn't just apply the new LoRAs - it extends the lora_stack it receives and passes the whole thing on. So the metadata chain stays intact: your sampler still knows about every single LoRA that touched the model, not just the ones this node added.

How it works

The node takes your already-loaded model and clip, plus an optional incoming lora_stack, and five pairs of LoraName/LoraWeight slots. For each LoRA you pick it calls the same load_lora_for_models machinery the stock LoraLoader uses - there's no custom math here, just a convenient wrapper. Then it appends each applied LoRA to the incoming stack as (name, weight, weight) tuples and returns the grown stack alongside the modified model and clip.

The weight being stored twice in each tuple isn't a quirk - it's how ComfyUI represents a LoRA applied at the same strength to both the model and the CLIP. When the Mango samplers log Lora hashes, they read this stack, so the saved metadata reflects the real full LoRA history of the image.

Inputs and outputs

The inputs that matter:

• model / clip - wire these from your loader (any loader, not just Mango's).
• lora_stack - optional; the stack from a Mango loader if you want to keep the chain growing. Feed it in to extend it, or leave it empty to start fresh.
• LoraName1…LoraName5 / LoraWeight1…LoraWeight5 - the five new LoRA slots. None skips a slot; weights default to 1.0, range −10 to 10.

Outputs are model, clip, and the extended lora_stack, which all wire straight into your sampler.

The order of operations is worth a moment: LoRAs from an incoming stack were applied by whatever node built it, then the five slots here apply in slot order. If you care about which LoRA "wins" a conflicting attribute, remember that later-applied LoRAs in ComfyUI generally take precedence - so put your priority LoRA in the highest slot.

Installing it

This node ships in the Mango Node Pack, so it arrives with the rest of the family:

cd ComfyUI/custom_nodes
git clone https://github.com/mang01010/MangoNodePack

Restart ComfyUI, or use ComfyUI Manager and search "Mango Node Pack." The README's only extra dependency is safetensors, already present in a standard install.

Gotchas

The honest take: if you're stacking this many LoRAs, you're in diminishing-returns territory. Ten LoRAs at 1.0 each rarely looks better than five well-chosen ones, and the pack README itself warns that some LoRA hashes won't be recognized by Civitai (usually because the model was re-uploaded/updated - re-download the LoRA and the hash matches again). Also, this node expects a lora_stack input even if you leave it empty - a bare workflow that starts from a stock Checkpoint Loader and feeds no stack works fine, but you won't get stack-based metadata logging unless something upstream actually produced a stack. For the "more LoRAs, same paper trail" use case, this is the one to reach for.